Tag Archives: Security

Advancing Protection in Chrome on Android

Posted by David Adrian, Javier Castro & Peter Kotwicz, Chrome Security Team

Android recently announced Advanced Protection, which extends Google’s Advanced Protection Program to a device-level security setting for Android users that need heightened security—such as journalists, elected officials, and public figures. Advanced Protection gives you the ability to activate Google’s strongest security for mobile devices, providing greater peace of mind that you’re better protected against the most sophisticated threats.

Advanced Protection acts as a single control point for at-risk users on Android that enables important security settings across applications, including many of your favorite Google apps, including Chrome. In this post, we’d like to do a deep dive into the Chrome features that are integrated with Advanced Protection, and how enterprises and users outside of Advanced Protection can leverage them.

Android Advanced Protection integrates with Chrome on Android in three main ways:

  • Enables the “Always Use Secure Connections” setting for both public and private sites, so that users are protected from attackers reading confidential data or injecting malicious content into insecure plaintext HTTP connections. Insecure HTTP represents less than 1% of page loads for Chrome on Android.
  • Enables full Site Isolation on mobile devices with 4GB+ RAM, so that potentially malicious sites are never loaded in the same process as legitimate websites. Desktop Chrome clients already have full Site Isolation.
  • Reduces attack surface by disabling Javascript optimizations, so that Chrome has a smaller attack surface and is harder to exploit.

Let’s take a look at all three, learn what they do, and how they can be controlled outside of Advanced Protection.

Always Use Secure Connections

“Always Use Secure Connections” (also known as HTTPS-First Mode in blog posts and HTTPS-Only Mode in the enterprise policy) is a Chrome setting that forces HTTPS wherever possible, and asks for explicit permission from you before connecting to a site insecurely. There may be attackers attempting to interpose on connections on any network, whether that network is a coffee shop, airport, or an Internet backbone. This setting protects users from these attackers reading confidential data and injecting malicious content into otherwise innocuous webpages. This is particularly useful for Advanced Protection users, since in 2023, plaintext HTTP was used as an exploitation vector during the Egyptian election.

Beyond Advanced Protection, we previously posted about how our goal is to eventually enable “Always Use Secure Connections” by default for all Chrome users. As we work towards this goal, in the last two years we have quietly been enabling it in more places beyond Advanced Protection, to help protect more users in risky situations, while limiting the number of warnings users might click through:

  • We added a new variant of the setting that only warns on public sites, and doesn’t warn on local networks or single-label hostnames (e.g. 192.168.0.1, shortlink/, 10.0.0.1). These names often cannot be issued a publicly-trusted HTTPS certificate. This variant protects against most threats—accessing a public website insecurely—but still allows for users to access local sites, which may be on a more trusted network, without seeing a warning.
  • We’ve automatically enabled “Always Use Secure Connections” for public sites in Incognito Mode for the last year, since Chrome 127 in June 2024.
  • We automatically prevent downgrades from HTTPS to plaintext HTTP on sites that Chrome knows you typically access over HTTPS (a heuristic version of the HSTS header), since Chrome 133 in January 2025.

Always Use Secure Connections has two modes—warn on insecure public sites, and warn on any insecure site.

Any user can enable “Always Use Secure Connections” in the Chrome Privacy and Security settings, regardless of if they’re using Advanced Protection. Users can choose if they would like to warn on any insecure site, or only insecure public sites. Enterprises can opt their fleet into either mode, and set exceptions using the HTTPSOnlyMode and HTTPAllowlist policies, respectively. Website operators should protect their users' confidentiality, ensure their content is delivered exactly as they intended, and avoid warnings, by deploying HTTPS.

Full Site Isolation

Site Isolation is a security feature in Chrome that isolates each website into its own rendering OS process. This means that different websites, even if loaded in a single tab of the same browser window, are kept completely separate from each other in memory. This isolation prevents a malicious website from accessing data or code from another website, even if that malicious website manages to exploit a vulnerability in Chrome’s renderer—a second bug to escape the renderer sandbox is required to access other sites. Site isolation improves security, but requires extra memory to have one process per site. Chrome Desktop isolates all sites by default. However, Android is particularly sensitive to memory usage, so for mobile Android form factors, when Advanced Protection is off, Chrome will only isolate a site if a user logs into that site, or if the user submits a form on that site. On Android devices with 4GB+ RAM in Advanced Protection (and on all desktop clients), Chrome will isolate all sites. Full Site Isolation significantly reduces the risk of cross-site data leakage for Advanced Protection users.

JavaScript Optimizations and Security

Advanced Protection reduces the attack surface of Chrome by disabling the higher-level optimizing Javascript compilers inside V8. V8 is Chrome’s high-performance Javascript and WebAssembly engine. The optimizing compilers in V8 make certain websites run faster, however they historically also have been a source of known exploitation of Chrome. Of all the patched security bugs in V8 with known exploitation, disabling the optimizers would have mitigated ~50%. However, the optimizers are why Chrome scores the highest on industry-wide benchmarks such as Speedometer. Disabling the optimizers blocks a large class of exploits, at the cost of causing performance issues for some websites.

Javascript optimizers can be disabled outside of Advanced Protection Mode via the “Javascript optimization & security” Site Setting. The Site Setting also enables users to disable/enable Javascript optimizers on a per-site basis. Disabling these optimizing compilers is not limited to Advanced Protection. Since Chrome 133, we’ve exposed this as a Site Setting that allows users to enable or disable the higher-level optimizing compilers on a per-site basis, as well as change the default.

Settings -> Privacy and Security -> Javascript optimization and security

This setting can be controlled by the DefaultJavaScriptOptimizerSetting enterprise policy, alongside JavaScriptOptimizerAllowedForSites and JavaScriptOptimizerBlockedForSites for managing the allowlist and denylist. Enterprises can use this policy to block access to the optimizer, while still allowlisting1 the SaaS vendors their employees use on a daily basis. It’s available on Android and desktop platforms

Chrome aims for the default configuration to be secure for all its users, and we’re continuing to raise the bar for V8 security in the default configuration by rolling out the V8 sandbox.

Protecting All Users

Billions of people use Chrome and Android, and not all of them have the same risk profile. Less sophisticated attacks by commodity malware can be very lucrative for attackers when done at scale, but so can sophisticated attacks on targeted users. This means that we cannot expect the security tradeoffs we make for the default configuration of Chrome to be suitable for everyone.

Advanced Protection, and the security settings associated with it, are a way for users with varying risk profiles to tailor Chrome to their security needs, either as an individual at-risk user. Enterprises with a fleet of managed Chrome installations can also enable the underlying settings now. Advanced Protection is available on Android 16 in Chrome 137+.

We additionally recommend at-risk users join the Advanced Protection Program with their Google accounts, which will require the account to use phishing-resistant multi-factor authentication methods and enable Advanced Protection on any of the user’s Android devices. We also recommend users enable automatic updates and always keep their Android phones and web browsers up to date.

Notes

  1. Allowlisting only works on platforms capable of full site isolation—any desktop platform and Android devices with 2GB+ RAM. This is because internally allowlisting is dependent on origin isolation. ↩

Mitigating prompt injection attacks with a layered defense strategy

Posted by Google GenAI Security Team

With the rapid adoption of generative AI, a new wave of threats is emerging across the industry with the aim of manipulating the AI systems themselves. One such emerging attack vector is indirect prompt injections. Unlike direct prompt injections, where an attacker directly inputs malicious commands into a prompt, indirect prompt injections involve hidden malicious instructions within external data sources. These may include emails, documents, or calendar invites that instruct AI to exfiltrate user data or execute other rogue actions. As more governments, businesses, and individuals adopt generative AI to get more done, this subtle yet potentially potent attack becomes increasingly pertinent across the industry, demanding immediate attention and robust security measures.


At Google, our teams have a longstanding precedent of investing in a defense-in-depth strategy, including robust evaluation, threat analysis, AI security best practices, AI red-teaming, adversarial training, and model hardening for generative AI tools. This approach enables safer adoption of Gemini in Google Workspace and the Gemini app (we refer to both in this blog as “Gemini” for simplicity). Below we describe our prompt injection mitigation product strategy based on extensive research, development, and deployment of improved security mitigations.


A layered security approach

Google has taken a layered security approach introducing security measures designed for each stage of the prompt lifecycle. From Gemini 2.5 model hardening, to purpose-built machine learning (ML) models detecting malicious instructions, to system-level safeguards, we are meaningfully elevating the difficulty, expense, and complexity faced by an attacker. This approach compels adversaries to resort to methods that are either more easily identified or demand greater resources. 


Our model training with adversarial data significantly enhanced our defenses against indirect prompt injection attacks in Gemini 2.5 models (technical details). This inherent model resilience is augmented with additional defenses that we built directly into Gemini, including: 


  1. Prompt injection content classifiers

  2. Security thought reinforcement

  3. Markdown sanitization and suspicious URL redaction

  4. User confirmation framework

  5. End-user security mitigation notifications


This layered approach to our security strategy strengthens the overall security framework for Gemini – throughout the prompt lifecycle and across diverse attack techniques.


1. Prompt injection content classifiers


Through collaboration with leading AI security researchers via Google's AI Vulnerability Reward Program (VRP), we've curated one of the world’s most advanced catalogs of generative AI vulnerabilities and adversarial data. Utilizing this resource, we built and are in the process of rolling out proprietary machine learning models that can detect malicious prompts and instructions within various formats, such as emails and files, drawing from real-world examples. Consequently, when users query Workspace data with Gemini, the content classifiers filter out harmful data containing malicious instructions, helping to ensure a secure end-to-end user experience by retaining only safe content. For example, if a user receives an email in Gmail that includes malicious instructions, our content classifiers help to detect and disregard malicious instructions, then generate a safe response for the user. This is in addition to built-in defenses in Gmail that automatically block more than 99.9% of spam, phishing attempts, and malware.


A diagram of Gemini’s actions based on the detection of the malicious instructions by content classifiers.


2. Security thought reinforcement


This technique adds targeted security instructions surrounding the prompt content to remind the large language model (LLM) to perform the user-directed task and ignore any adversarial instructions that could be present in the content. With this approach, we steer the LLM to stay focused on the task and ignore harmful or malicious requests added by a threat actor to execute indirect prompt injection attacks.

A diagram of Gemini’s actions based on additional protection provided by the security thought reinforcement technique. 


3. Markdown sanitization and suspicious URL redaction 


Our markdown sanitizer identifies external image URLs and will not render them, making the “EchoLeak” 0-click image rendering exfiltration vulnerability not applicable to Gemini. From there, a key protection against prompt injection and data exfiltration attacks occurs at the URL level. With external data containing dynamic URLs, users may encounter unknown risks as these URLs may be designed for indirect prompt injections and data exfiltration attacks. Malicious instructions executed on a user's behalf may also generate harmful URLs. With Gemini, our defense system includes suspicious URL detection based on Google Safe Browsing to differentiate between safe and unsafe links, providing a secure experience by helping to prevent URL-based attacks. For example, if a document contains malicious URLs and a user is summarizing the content with Gemini, the suspicious URLs will be redacted in Gemini’s response. 


Gemini in Gmail provides a summary of an email thread. In the summary, there is an unsafe URL. That URL is redacted in the response and is replaced with the text “suspicious link removed”. 


4. User confirmation framework


Gemini also features a contextual user confirmation system. This framework enables Gemini to require user confirmation for certain actions, also known as “Human-In-The-Loop” (HITL), using these responses to bolster security and streamline the user experience. For example, potentially risky operations like deleting a calendar event may trigger an explicit user confirmation request, thereby helping to prevent undetected or immediate execution of the operation.


The Gemini app with instructions to delete all events on Saturday. Gemini responds with the events found on Google Calendar and asks the user to confirm this action.


5. End-user security mitigation notifications


A key aspect to keeping our users safe is sharing details on attacks that we’ve stopped so users can watch out for similar attacks in the future. To that end, when security issues are mitigated with our built-in defenses, end users are provided with contextual information allowing them to learn more via dedicated help center articles. For example, if Gemini summarizes a file containing malicious instructions and one of Google’s prompt injection defenses mitigates the situation, a security notification with a “Learn more” link will be displayed for the user. Users are encouraged to become more familiar with our prompt injection defenses by reading the Help Center article. 


Gemini in Docs with instructions to provide a summary of a file. Suspicious content was detected and a response was not provided. There is a yellow security notification banner for the user and a statement that Gemini’s response has been removed, with a “Learn more” link to a relevant Help Center article.

Moving forward


Our comprehensive prompt injection security strategy strengthens the overall security framework for Gemini. Beyond the techniques described above, it also involves rigorous testing through manual and automated red teams, generative AI security BugSWAT events, strong security standards like our Secure AI Framework (SAIF), and partnerships with both external researchers via the Google AI Vulnerability Reward Program (VRP) and industry peers via the Coalition for Secure AI (CoSAI). Our commitment to trust includes collaboration with the security community to responsibly disclose AI security vulnerabilities, share our latest threat intelligence on ways we see bad actors trying to leverage AI, and offering insights into our work to build stronger prompt injection defenses. 


Working closely with industry partners is crucial to building stronger protections for all of our users. To that end, we’re fortunate to have strong collaborative partnerships with numerous researchers, such as Ben Nassi (Confidentiality), Stav Cohen (Technion), and Or Yair (SafeBreach), as well as other AI Security researchers participating in our BugSWAT events and AI VRP program. We appreciate the work of these researchers and others in the community to help us red team and refine our defenses.


We continue working to make upcoming Gemini models inherently more resilient and add additional prompt injection defenses directly into Gemini later this year. To learn more about Google’s progress and research on generative AI threat actors, attack techniques, and vulnerabilities, take a look at the following resources:


Mitigating prompt injection attacks with a layered defense strategy

Posted by Google GenAI Security Team

With the rapid adoption of generative AI, a new wave of threats is emerging across the industry with the aim of manipulating the AI systems themselves. One such emerging attack vector is indirect prompt injections. Unlike direct prompt injections, where an attacker directly inputs malicious commands into a prompt, indirect prompt injections involve hidden malicious instructions within external data sources. These may include emails, documents, or calendar invites that instruct AI to exfiltrate user data or execute other rogue actions. As more governments, businesses, and individuals adopt generative AI to get more done, this subtle yet potentially potent attack becomes increasingly pertinent across the industry, demanding immediate attention and robust security measures.


At Google, our teams have a longstanding precedent of investing in a defense-in-depth strategy, including robust evaluation, threat analysis, AI security best practices, AI red-teaming, adversarial training, and model hardening for generative AI tools. This approach enables safer adoption of Gemini in Google Workspace and the Gemini app (we refer to both in this blog as “Gemini” for simplicity). Below we describe our prompt injection mitigation product strategy based on extensive research, development, and deployment of improved security mitigations.


A layered security approach

Google has taken a layered security approach introducing security measures designed for each stage of the prompt lifecycle. From Gemini 2.5 model hardening, to purpose-built machine learning (ML) models detecting malicious instructions, to system-level safeguards, we are meaningfully elevating the difficulty, expense, and complexity faced by an attacker. This approach compels adversaries to resort to methods that are either more easily identified or demand greater resources. 


Our model training with adversarial data significantly enhanced our defenses against indirect prompt injection attacks in Gemini 2.5 models (technical details). This inherent model resilience is augmented with additional defenses that we built directly into Gemini, including: 


  1. Prompt injection content classifiers

  2. Security thought reinforcement

  3. Markdown sanitization and suspicious URL redaction

  4. User confirmation framework

  5. End-user security mitigation notifications


This layered approach to our security strategy strengthens the overall security framework for Gemini – throughout the prompt lifecycle and across diverse attack techniques.


1. Prompt injection content classifiers


Through collaboration with leading AI security researchers via Google's AI Vulnerability Reward Program (VRP), we've curated one of the world’s most advanced catalogs of generative AI vulnerabilities and adversarial data. Utilizing this resource, we built and are in the process of rolling out proprietary machine learning models that can detect malicious prompts and instructions within various formats, such as emails and files, drawing from real-world examples. Consequently, when users query Workspace data with Gemini, the content classifiers filter out harmful data containing malicious instructions, helping to ensure a secure end-to-end user experience by retaining only safe content. For example, if a user receives an email in Gmail that includes malicious instructions, our content classifiers help to detect and disregard malicious instructions, then generate a safe response for the user. This is in addition to built-in defenses in Gmail that automatically block more than 99.9% of spam, phishing attempts, and malware.


A diagram of Gemini’s actions based on the detection of the malicious instructions by content classifiers.


2. Security thought reinforcement


This technique adds targeted security instructions surrounding the prompt content to remind the large language model (LLM) to perform the user-directed task and ignore any adversarial instructions that could be present in the content. With this approach, we steer the LLM to stay focused on the task and ignore harmful or malicious requests added by a threat actor to execute indirect prompt injection attacks.

A diagram of Gemini’s actions based on additional protection provided by the security thought reinforcement technique. 


3. Markdown sanitization and suspicious URL redaction 


Our markdown sanitizer identifies external image URLs and will not render them, making the “EchoLeak” 0-click image rendering exfiltration vulnerability not applicable to Gemini. From there, a key protection against prompt injection and data exfiltration attacks occurs at the URL level. With external data containing dynamic URLs, users may encounter unknown risks as these URLs may be designed for indirect prompt injections and data exfiltration attacks. Malicious instructions executed on a user's behalf may also generate harmful URLs. With Gemini, our defense system includes suspicious URL detection based on Google Safe Browsing to differentiate between safe and unsafe links, providing a secure experience by helping to prevent URL-based attacks. For example, if a document contains malicious URLs and a user is summarizing the content with Gemini, the suspicious URLs will be redacted in Gemini’s response. 


Gemini in Gmail provides a summary of an email thread. In the summary, there is an unsafe URL. That URL is redacted in the response and is replaced with the text “suspicious link removed”. 


4. User confirmation framework


Gemini also features a contextual user confirmation system. This framework enables Gemini to require user confirmation for certain actions, also known as “Human-In-The-Loop” (HITL), using these responses to bolster security and streamline the user experience. For example, potentially risky operations like deleting a calendar event may trigger an explicit user confirmation request, thereby helping to prevent undetected or immediate execution of the operation.


The Gemini app with instructions to delete all events on Saturday. Gemini responds with the events found on Google Calendar and asks the user to confirm this action.


5. End-user security mitigation notifications


A key aspect to keeping our users safe is sharing details on attacks that we’ve stopped so users can watch out for similar attacks in the future. To that end, when security issues are mitigated with our built-in defenses, end users are provided with contextual information allowing them to learn more via dedicated help center articles. For example, if Gemini summarizes a file containing malicious instructions and one of Google’s prompt injection defenses mitigates the situation, a security notification with a “Learn more” link will be displayed for the user. Users are encouraged to become more familiar with our prompt injection defenses by reading the Help Center article. 


Gemini in Docs with instructions to provide a summary of a file. Suspicious content was detected and a response was not provided. There is a yellow security notification banner for the user and a statement that Gemini’s response has been removed, with a “Learn more” link to a relevant Help Center article.

Moving forward


Our comprehensive prompt injection security strategy strengthens the overall security framework for Gemini. Beyond the techniques described above, it also involves rigorous testing through manual and automated red teams, generative AI security BugSWAT events, strong security standards like our Secure AI Framework (SAIF), and partnerships with both external researchers via the Google AI Vulnerability Reward Program (VRP) and industry peers via the Coalition for Secure AI (CoSAI). Our commitment to trust includes collaboration with the security community to responsibly disclose AI security vulnerabilities, share our latest threat intelligence on ways we see bad actors trying to leverage AI, and offering insights into our work to build stronger prompt injection defenses. 


Working closely with industry partners is crucial to building stronger protections for all of our users. To that end, we’re fortunate to have strong collaborative partnerships with numerous researchers, such as Ben Nassi (Confidentiality), Stav Cohen (Technion), and Or Yair (SafeBreach), as well as other AI Security researchers participating in our BugSWAT events and AI VRP program. We appreciate the work of these researchers and others in the community to help us red team and refine our defenses.


We continue working to make upcoming Gemini models inherently more resilient and add additional prompt injection defenses directly into Gemini later this year. To learn more about Google’s progress and research on generative AI threat actors, attack techniques, and vulnerabilities, take a look at the following resources:


Mitigating prompt injection attacks with a layered defense strategy

Posted by Google GenAI Security Team

With the rapid adoption of generative AI, a new wave of threats is emerging across the industry with the aim of manipulating the AI systems themselves. One such emerging attack vector is indirect prompt injections. Unlike direct prompt injections, where an attacker directly inputs malicious commands into a prompt, indirect prompt injections involve hidden malicious instructions within external data sources. These may include emails, documents, or calendar invites that instruct AI to exfiltrate user data or execute other rogue actions. As more governments, businesses, and individuals adopt generative AI to get more done, this subtle yet potentially potent attack becomes increasingly pertinent across the industry, demanding immediate attention and robust security measures.


At Google, our teams have a longstanding precedent of investing in a defense-in-depth strategy, including robust evaluation, threat analysis, AI security best practices, AI red-teaming, adversarial training, and model hardening for generative AI tools. This approach enables safer adoption of Gemini in Google Workspace and the Gemini app (we refer to both in this blog as “Gemini” for simplicity). Below we describe our prompt injection mitigation product strategy based on extensive research, development, and deployment of improved security mitigations.


A layered security approach

Google has taken a layered security approach introducing security measures designed for each stage of the prompt lifecycle. From Gemini 2.5 model hardening, to purpose-built machine learning (ML) models detecting malicious instructions, to system-level safeguards, we are meaningfully elevating the difficulty, expense, and complexity faced by an attacker. This approach compels adversaries to resort to methods that are either more easily identified or demand greater resources. 


Our model training with adversarial data significantly enhanced our defenses against indirect prompt injection attacks in Gemini 2.5 models (technical details). This inherent model resilience is augmented with additional defenses that we built directly into Gemini, including: 


  1. Prompt injection content classifiers

  2. Security thought reinforcement

  3. Markdown sanitization and suspicious URL redaction

  4. User confirmation framework

  5. End-user security mitigation notifications


This layered approach to our security strategy strengthens the overall security framework for Gemini – throughout the prompt lifecycle and across diverse attack techniques.


1. Prompt injection content classifiers


Through collaboration with leading AI security researchers via Google's AI Vulnerability Reward Program (VRP), we've curated one of the world’s most advanced catalogs of generative AI vulnerabilities and adversarial data. Utilizing this resource, we built and are in the process of rolling out proprietary machine learning models that can detect malicious prompts and instructions within various formats, such as emails and files, drawing from real-world examples. Consequently, when users query Workspace data with Gemini, the content classifiers filter out harmful data containing malicious instructions, helping to ensure a secure end-to-end user experience by retaining only safe content. For example, if a user receives an email in Gmail that includes malicious instructions, our content classifiers help to detect and disregard malicious instructions, then generate a safe response for the user. This is in addition to built-in defenses in Gmail that automatically block more than 99.9% of spam, phishing attempts, and malware.


A diagram of Gemini’s actions based on the detection of the malicious instructions by content classifiers.


2. Security thought reinforcement


This technique adds targeted security instructions surrounding the prompt content to remind the large language model (LLM) to perform the user-directed task and ignore any adversarial instructions that could be present in the content. With this approach, we steer the LLM to stay focused on the task and ignore harmful or malicious requests added by a threat actor to execute indirect prompt injection attacks.

A diagram of Gemini’s actions based on additional protection provided by the security thought reinforcement technique. 


3. Markdown sanitization and suspicious URL redaction 


Our markdown sanitizer identifies external image URLs and will not render them, making the “EchoLeak” 0-click image rendering exfiltration vulnerability not applicable to Gemini. From there, a key protection against prompt injection and data exfiltration attacks occurs at the URL level. With external data containing dynamic URLs, users may encounter unknown risks as these URLs may be designed for indirect prompt injections and data exfiltration attacks. Malicious instructions executed on a user's behalf may also generate harmful URLs. With Gemini, our defense system includes suspicious URL detection based on Google Safe Browsing to differentiate between safe and unsafe links, providing a secure experience by helping to prevent URL-based attacks. For example, if a document contains malicious URLs and a user is summarizing the content with Gemini, the suspicious URLs will be redacted in Gemini’s response. 


Gemini in Gmail provides a summary of an email thread. In the summary, there is an unsafe URL. That URL is redacted in the response and is replaced with the text “suspicious link removed”. 


4. User confirmation framework


Gemini also features a contextual user confirmation system. This framework enables Gemini to require user confirmation for certain actions, also known as “Human-In-The-Loop” (HITL), using these responses to bolster security and streamline the user experience. For example, potentially risky operations like deleting a calendar event may trigger an explicit user confirmation request, thereby helping to prevent undetected or immediate execution of the operation.


The Gemini app with instructions to delete all events on Saturday. Gemini responds with the events found on Google Calendar and asks the user to confirm this action.


5. End-user security mitigation notifications


A key aspect to keeping our users safe is sharing details on attacks that we’ve stopped so users can watch out for similar attacks in the future. To that end, when security issues are mitigated with our built-in defenses, end users are provided with contextual information allowing them to learn more via dedicated help center articles. For example, if Gemini summarizes a file containing malicious instructions and one of Google’s prompt injection defenses mitigates the situation, a security notification with a “Learn more” link will be displayed for the user. Users are encouraged to become more familiar with our prompt injection defenses by reading the Help Center article. 


Gemini in Docs with instructions to provide a summary of a file. Suspicious content was detected and a response was not provided. There is a yellow security notification banner for the user and a statement that Gemini’s response has been removed, with a “Learn more” link to a relevant Help Center article.

Moving forward


Our comprehensive prompt injection security strategy strengthens the overall security framework for Gemini. Beyond the techniques described above, it also involves rigorous testing through manual and automated red teams, generative AI security BugSWAT events, strong security standards like our Secure AI Framework (SAIF), and partnerships with both external researchers via the Google AI Vulnerability Reward Program (VRP) and industry peers via the Coalition for Secure AI (CoSAI). Our commitment to trust includes collaboration with the security community to responsibly disclose AI security vulnerabilities, share our latest threat intelligence on ways we see bad actors trying to leverage AI, and offering insights into our work to build stronger prompt injection defenses. 


Working closely with industry partners is crucial to building stronger protections for all of our users. To that end, we’re fortunate to have strong collaborative partnerships with numerous researchers, such as Ben Nassi (Confidentiality), Stav Cohen (Technion), and Or Yair (SafeBreach), as well as other AI Security researchers participating in our BugSWAT events and AI VRP program. We appreciate the work of these researchers and others in the community to help us red team and refine our defenses.


We continue working to make upcoming Gemini models inherently more resilient and add additional prompt injection defenses directly into Gemini later this year. To learn more about Google’s progress and research on generative AI threat actors, attack techniques, and vulnerabilities, take a look at the following resources:


Mitigating prompt injection attacks with a layered defense strategy

Posted by Google GenAI Security Team

With the rapid adoption of generative AI, a new wave of threats is emerging across the industry with the aim of manipulating the AI systems themselves. One such emerging attack vector is indirect prompt injections. Unlike direct prompt injections, where an attacker directly inputs malicious commands into a prompt, indirect prompt injections involve hidden malicious instructions within external data sources. These may include emails, documents, or calendar invites that instruct AI to exfiltrate user data or execute other rogue actions. As more governments, businesses, and individuals adopt generative AI to get more done, this subtle yet potentially potent attack becomes increasingly pertinent across the industry, demanding immediate attention and robust security measures.


At Google, our teams have a longstanding precedent of investing in a defense-in-depth strategy, including robust evaluation, threat analysis, AI security best practices, AI red-teaming, adversarial training, and model hardening for generative AI tools. This approach enables safer adoption of Gemini in Google Workspace and the Gemini app (we refer to both in this blog as “Gemini” for simplicity). Below we describe our prompt injection mitigation product strategy based on extensive research, development, and deployment of improved security mitigations.


A layered security approach

Google has taken a layered security approach introducing security measures designed for each stage of the prompt lifecycle. From Gemini 2.5 model hardening, to purpose-built machine learning (ML) models detecting malicious instructions, to system-level safeguards, we are meaningfully elevating the difficulty, expense, and complexity faced by an attacker. This approach compels adversaries to resort to methods that are either more easily identified or demand greater resources. 


Our model training with adversarial data significantly enhanced our defenses against indirect prompt injection attacks in Gemini 2.5 models (technical details). This inherent model resilience is augmented with additional defenses that we built directly into Gemini, including: 


  1. Prompt injection content classifiers

  2. Security thought reinforcement

  3. Markdown sanitization and suspicious URL redaction

  4. User confirmation framework

  5. End-user security mitigation notifications


This layered approach to our security strategy strengthens the overall security framework for Gemini – throughout the prompt lifecycle and across diverse attack techniques.


1. Prompt injection content classifiers


Through collaboration with leading AI security researchers via Google's AI Vulnerability Reward Program (VRP), we've curated one of the world’s most advanced catalogs of generative AI vulnerabilities and adversarial data. Utilizing this resource, we built and are in the process of rolling out proprietary machine learning models that can detect malicious prompts and instructions within various formats, such as emails and files, drawing from real-world examples. Consequently, when users query Workspace data with Gemini, the content classifiers filter out harmful data containing malicious instructions, helping to ensure a secure end-to-end user experience by retaining only safe content. For example, if a user receives an email in Gmail that includes malicious instructions, our content classifiers help to detect and disregard malicious instructions, then generate a safe response for the user. This is in addition to built-in defenses in Gmail that automatically block more than 99.9% of spam, phishing attempts, and malware.


A diagram of Gemini’s actions based on the detection of the malicious instructions by content classifiers.


2. Security thought reinforcement


This technique adds targeted security instructions surrounding the prompt content to remind the large language model (LLM) to perform the user-directed task and ignore any adversarial instructions that could be present in the content. With this approach, we steer the LLM to stay focused on the task and ignore harmful or malicious requests added by a threat actor to execute indirect prompt injection attacks.

A diagram of Gemini’s actions based on additional protection provided by the security thought reinforcement technique. 


3. Markdown sanitization and suspicious URL redaction 


Our markdown sanitizer identifies external image URLs and will not render them, making the “EchoLeak” 0-click image rendering exfiltration vulnerability not applicable to Gemini. From there, a key protection against prompt injection and data exfiltration attacks occurs at the URL level. With external data containing dynamic URLs, users may encounter unknown risks as these URLs may be designed for indirect prompt injections and data exfiltration attacks. Malicious instructions executed on a user's behalf may also generate harmful URLs. With Gemini, our defense system includes suspicious URL detection based on Google Safe Browsing to differentiate between safe and unsafe links, providing a secure experience by helping to prevent URL-based attacks. For example, if a document contains malicious URLs and a user is summarizing the content with Gemini, the suspicious URLs will be redacted in Gemini’s response. 


Gemini in Gmail provides a summary of an email thread. In the summary, there is an unsafe URL. That URL is redacted in the response and is replaced with the text “suspicious link removed”. 


4. User confirmation framework


Gemini also features a contextual user confirmation system. This framework enables Gemini to require user confirmation for certain actions, also known as “Human-In-The-Loop” (HITL), using these responses to bolster security and streamline the user experience. For example, potentially risky operations like deleting a calendar event may trigger an explicit user confirmation request, thereby helping to prevent undetected or immediate execution of the operation.


The Gemini app with instructions to delete all events on Saturday. Gemini responds with the events found on Google Calendar and asks the user to confirm this action.


5. End-user security mitigation notifications


A key aspect to keeping our users safe is sharing details on attacks that we’ve stopped so users can watch out for similar attacks in the future. To that end, when security issues are mitigated with our built-in defenses, end users are provided with contextual information allowing them to learn more via dedicated help center articles. For example, if Gemini summarizes a file containing malicious instructions and one of Google’s prompt injection defenses mitigates the situation, a security notification with a “Learn more” link will be displayed for the user. Users are encouraged to become more familiar with our prompt injection defenses by reading the Help Center article. 


Gemini in Docs with instructions to provide a summary of a file. Suspicious content was detected and a response was not provided. There is a yellow security notification banner for the user and a statement that Gemini’s response has been removed, with a “Learn more” link to a relevant Help Center article.

Moving forward


Our comprehensive prompt injection security strategy strengthens the overall security framework for Gemini. Beyond the techniques described above, it also involves rigorous testing through manual and automated red teams, generative AI security BugSWAT events, strong security standards like our Secure AI Framework (SAIF), and partnerships with both external researchers via the Google AI Vulnerability Reward Program (VRP) and industry peers via the Coalition for Secure AI (CoSAI). Our commitment to trust includes collaboration with the security community to responsibly disclose AI security vulnerabilities, share our latest threat intelligence on ways we see bad actors trying to leverage AI, and offering insights into our work to build stronger prompt injection defenses. 


Working closely with industry partners is crucial to building stronger protections for all of our users. To that end, we’re fortunate to have strong collaborative partnerships with numerous researchers, such as Ben Nassi (Confidentiality), Stav Cohen (Technion), and Or Yair (SafeBreach), as well as other AI Security researchers participating in our BugSWAT events and AI VRP program. We appreciate the work of these researchers and others in the community to help us red team and refine our defenses.


We continue working to make upcoming Gemini models inherently more resilient and add additional prompt injection defenses directly into Gemini later this year. To learn more about Google’s progress and research on generative AI threat actors, attack techniques, and vulnerabilities, take a look at the following resources:


Mitigating prompt injection attacks with a layered defense strategy

Posted by Google GenAI Security Team

With the rapid adoption of generative AI, a new wave of threats is emerging across the industry with the aim of manipulating the AI systems themselves. One such emerging attack vector is indirect prompt injections. Unlike direct prompt injections, where an attacker directly inputs malicious commands into a prompt, indirect prompt injections involve hidden malicious instructions within external data sources. These may include emails, documents, or calendar invites that instruct AI to exfiltrate user data or execute other rogue actions. As more governments, businesses, and individuals adopt generative AI to get more done, this subtle yet potentially potent attack becomes increasingly pertinent across the industry, demanding immediate attention and robust security measures.


At Google, our teams have a longstanding precedent of investing in a defense-in-depth strategy, including robust evaluation, threat analysis, AI security best practices, AI red-teaming, adversarial training, and model hardening for generative AI tools. This approach enables safer adoption of Gemini in Google Workspace and the Gemini app (we refer to both in this blog as “Gemini” for simplicity). Below we describe our prompt injection mitigation product strategy based on extensive research, development, and deployment of improved security mitigations.


A layered security approach

Google has taken a layered security approach introducing security measures designed for each stage of the prompt lifecycle. From Gemini 2.5 model hardening, to purpose-built machine learning (ML) models detecting malicious instructions, to system-level safeguards, we are meaningfully elevating the difficulty, expense, and complexity faced by an attacker. This approach compels adversaries to resort to methods that are either more easily identified or demand greater resources. 


Our model training with adversarial data significantly enhanced our defenses against indirect prompt injection attacks in Gemini 2.5 models (technical details). This inherent model resilience is augmented with additional defenses that we built directly into Gemini, including: 


  1. Prompt injection content classifiers

  2. Security thought reinforcement

  3. Markdown sanitization and suspicious URL redaction

  4. User confirmation framework

  5. End-user security mitigation notifications


This layered approach to our security strategy strengthens the overall security framework for Gemini – throughout the prompt lifecycle and across diverse attack techniques.


1. Prompt injection content classifiers


Through collaboration with leading AI security researchers via Google's AI Vulnerability Reward Program (VRP), we've curated one of the world’s most advanced catalogs of generative AI vulnerabilities and adversarial data. Utilizing this resource, we built and are in the process of rolling out proprietary machine learning models that can detect malicious prompts and instructions within various formats, such as emails and files, drawing from real-world examples. Consequently, when users query Workspace data with Gemini, the content classifiers filter out harmful data containing malicious instructions, helping to ensure a secure end-to-end user experience by retaining only safe content. For example, if a user receives an email in Gmail that includes malicious instructions, our content classifiers help to detect and disregard malicious instructions, then generate a safe response for the user. This is in addition to built-in defenses in Gmail that automatically block more than 99.9% of spam, phishing attempts, and malware.


A diagram of Gemini’s actions based on the detection of the malicious instructions by content classifiers.


2. Security thought reinforcement


This technique adds targeted security instructions surrounding the prompt content to remind the large language model (LLM) to perform the user-directed task and ignore any adversarial instructions that could be present in the content. With this approach, we steer the LLM to stay focused on the task and ignore harmful or malicious requests added by a threat actor to execute indirect prompt injection attacks.

A diagram of Gemini’s actions based on additional protection provided by the security thought reinforcement technique. 


3. Markdown sanitization and suspicious URL redaction 


Our markdown sanitizer identifies external image URLs and will not render them, making the “EchoLeak” 0-click image rendering exfiltration vulnerability not applicable to Gemini. From there, a key protection against prompt injection and data exfiltration attacks occurs at the URL level. With external data containing dynamic URLs, users may encounter unknown risks as these URLs may be designed for indirect prompt injections and data exfiltration attacks. Malicious instructions executed on a user's behalf may also generate harmful URLs. With Gemini, our defense system includes suspicious URL detection based on Google Safe Browsing to differentiate between safe and unsafe links, providing a secure experience by helping to prevent URL-based attacks. For example, if a document contains malicious URLs and a user is summarizing the content with Gemini, the suspicious URLs will be redacted in Gemini’s response. 


Gemini in Gmail provides a summary of an email thread. In the summary, there is an unsafe URL. That URL is redacted in the response and is replaced with the text “suspicious link removed”. 


4. User confirmation framework


Gemini also features a contextual user confirmation system. This framework enables Gemini to require user confirmation for certain actions, also known as “Human-In-The-Loop” (HITL), using these responses to bolster security and streamline the user experience. For example, potentially risky operations like deleting a calendar event may trigger an explicit user confirmation request, thereby helping to prevent undetected or immediate execution of the operation.


The Gemini app with instructions to delete all events on Saturday. Gemini responds with the events found on Google Calendar and asks the user to confirm this action.


5. End-user security mitigation notifications


A key aspect to keeping our users safe is sharing details on attacks that we’ve stopped so users can watch out for similar attacks in the future. To that end, when security issues are mitigated with our built-in defenses, end users are provided with contextual information allowing them to learn more via dedicated help center articles. For example, if Gemini summarizes a file containing malicious instructions and one of Google’s prompt injection defenses mitigates the situation, a security notification with a “Learn more” link will be displayed for the user. Users are encouraged to become more familiar with our prompt injection defenses by reading the Help Center article. 


Gemini in Docs with instructions to provide a summary of a file. Suspicious content was detected and a response was not provided. There is a yellow security notification banner for the user and a statement that Gemini’s response has been removed, with a “Learn more” link to a relevant Help Center article.

Moving forward


Our comprehensive prompt injection security strategy strengthens the overall security framework for Gemini. Beyond the techniques described above, it also involves rigorous testing through manual and automated red teams, generative AI security BugSWAT events, strong security standards like our Secure AI Framework (SAIF), and partnerships with both external researchers via the Google AI Vulnerability Reward Program (VRP) and industry peers via the Coalition for Secure AI (CoSAI). Our commitment to trust includes collaboration with the security community to responsibly disclose AI security vulnerabilities, share our latest threat intelligence on ways we see bad actors trying to leverage AI, and offering insights into our work to build stronger prompt injection defenses. 


Working closely with industry partners is crucial to building stronger protections for all of our users. To that end, we’re fortunate to have strong collaborative partnerships with numerous researchers, such as Ben Nassi (Confidentiality), Stav Cohen (Technion), and Or Yair (SafeBreach), as well as other AI Security researchers participating in our BugSWAT events and AI VRP program. We appreciate the work of these researchers and others in the community to help us red team and refine our defenses.


We continue working to make upcoming Gemini models inherently more resilient and add additional prompt injection defenses directly into Gemini later this year. To learn more about Google’s progress and research on generative AI threat actors, attack techniques, and vulnerabilities, take a look at the following resources:


Mitigating prompt injection attacks with a layered defense strategy

Posted by Google GenAI Security Team

With the rapid adoption of generative AI, a new wave of threats is emerging across the industry with the aim of manipulating the AI systems themselves. One such emerging attack vector is indirect prompt injections. Unlike direct prompt injections, where an attacker directly inputs malicious commands into a prompt, indirect prompt injections involve hidden malicious instructions within external data sources. These may include emails, documents, or calendar invites that instruct AI to exfiltrate user data or execute other rogue actions. As more governments, businesses, and individuals adopt generative AI to get more done, this subtle yet potentially potent attack becomes increasingly pertinent across the industry, demanding immediate attention and robust security measures.


At Google, our teams have a longstanding precedent of investing in a defense-in-depth strategy, including robust evaluation, threat analysis, AI security best practices, AI red-teaming, adversarial training, and model hardening for generative AI tools. This approach enables safer adoption of Gemini in Google Workspace and the Gemini app (we refer to both in this blog as “Gemini” for simplicity). Below we describe our prompt injection mitigation product strategy based on extensive research, development, and deployment of improved security mitigations.


A layered security approach

Google has taken a layered security approach introducing security measures designed for each stage of the prompt lifecycle. From Gemini 2.5 model hardening, to purpose-built machine learning (ML) models detecting malicious instructions, to system-level safeguards, we are meaningfully elevating the difficulty, expense, and complexity faced by an attacker. This approach compels adversaries to resort to methods that are either more easily identified or demand greater resources. 


Our model training with adversarial data significantly enhanced our defenses against indirect prompt injection attacks in Gemini 2.5 models (technical details). This inherent model resilience is augmented with additional defenses that we built directly into Gemini, including: 


  1. Prompt injection content classifiers

  2. Security thought reinforcement

  3. Markdown sanitization and suspicious URL redaction

  4. User confirmation framework

  5. End-user security mitigation notifications


This layered approach to our security strategy strengthens the overall security framework for Gemini – throughout the prompt lifecycle and across diverse attack techniques.


1. Prompt injection content classifiers


Through collaboration with leading AI security researchers via Google's AI Vulnerability Reward Program (VRP), we've curated one of the world’s most advanced catalogs of generative AI vulnerabilities and adversarial data. Utilizing this resource, we built and are in the process of rolling out proprietary machine learning models that can detect malicious prompts and instructions within various formats, such as emails and files, drawing from real-world examples. Consequently, when users query Workspace data with Gemini, the content classifiers filter out harmful data containing malicious instructions, helping to ensure a secure end-to-end user experience by retaining only safe content. For example, if a user receives an email in Gmail that includes malicious instructions, our content classifiers help to detect and disregard malicious instructions, then generate a safe response for the user. This is in addition to built-in defenses in Gmail that automatically block more than 99.9% of spam, phishing attempts, and malware.


A diagram of Gemini’s actions based on the detection of the malicious instructions by content classifiers.


2. Security thought reinforcement


This technique adds targeted security instructions surrounding the prompt content to remind the large language model (LLM) to perform the user-directed task and ignore any adversarial instructions that could be present in the content. With this approach, we steer the LLM to stay focused on the task and ignore harmful or malicious requests added by a threat actor to execute indirect prompt injection attacks.

A diagram of Gemini’s actions based on additional protection provided by the security thought reinforcement technique. 


3. Markdown sanitization and suspicious URL redaction 


Our markdown sanitizer identifies external image URLs and will not render them, making the “EchoLeak” 0-click image rendering exfiltration vulnerability not applicable to Gemini. From there, a key protection against prompt injection and data exfiltration attacks occurs at the URL level. With external data containing dynamic URLs, users may encounter unknown risks as these URLs may be designed for indirect prompt injections and data exfiltration attacks. Malicious instructions executed on a user's behalf may also generate harmful URLs. With Gemini, our defense system includes suspicious URL detection based on Google Safe Browsing to differentiate between safe and unsafe links, providing a secure experience by helping to prevent URL-based attacks. For example, if a document contains malicious URLs and a user is summarizing the content with Gemini, the suspicious URLs will be redacted in Gemini’s response. 


Gemini in Gmail provides a summary of an email thread. In the summary, there is an unsafe URL. That URL is redacted in the response and is replaced with the text “suspicious link removed”. 


4. User confirmation framework


Gemini also features a contextual user confirmation system. This framework enables Gemini to require user confirmation for certain actions, also known as “Human-In-The-Loop” (HITL), using these responses to bolster security and streamline the user experience. For example, potentially risky operations like deleting a calendar event may trigger an explicit user confirmation request, thereby helping to prevent undetected or immediate execution of the operation.


The Gemini app with instructions to delete all events on Saturday. Gemini responds with the events found on Google Calendar and asks the user to confirm this action.


5. End-user security mitigation notifications


A key aspect to keeping our users safe is sharing details on attacks that we’ve stopped so users can watch out for similar attacks in the future. To that end, when security issues are mitigated with our built-in defenses, end users are provided with contextual information allowing them to learn more via dedicated help center articles. For example, if Gemini summarizes a file containing malicious instructions and one of Google’s prompt injection defenses mitigates the situation, a security notification with a “Learn more” link will be displayed for the user. Users are encouraged to become more familiar with our prompt injection defenses by reading the Help Center article. 


Gemini in Docs with instructions to provide a summary of a file. Suspicious content was detected and a response was not provided. There is a yellow security notification banner for the user and a statement that Gemini’s response has been removed, with a “Learn more” link to a relevant Help Center article.

Moving forward


Our comprehensive prompt injection security strategy strengthens the overall security framework for Gemini. Beyond the techniques described above, it also involves rigorous testing through manual and automated red teams, generative AI security BugSWAT events, strong security standards like our Secure AI Framework (SAIF), and partnerships with both external researchers via the Google AI Vulnerability Reward Program (VRP) and industry peers via the Coalition for Secure AI (CoSAI). Our commitment to trust includes collaboration with the security community to responsibly disclose AI security vulnerabilities, share our latest threat intelligence on ways we see bad actors trying to leverage AI, and offering insights into our work to build stronger prompt injection defenses. 


Working closely with industry partners is crucial to building stronger protections for all of our users. To that end, we’re fortunate to have strong collaborative partnerships with numerous researchers, such as Ben Nassi (Confidentiality), Stav Cohen (Technion), and Or Yair (SafeBreach), as well as other AI Security researchers participating in our BugSWAT events and AI VRP program. We appreciate the work of these researchers and others in the community to help us red team and refine our defenses.


We continue working to make upcoming Gemini models inherently more resilient and add additional prompt injection defenses directly into Gemini later this year. To learn more about Google’s progress and research on generative AI threat actors, attack techniques, and vulnerabilities, take a look at the following resources:


Mitigating prompt injection attacks with a layered defense strategy

Posted by Google GenAI Security Team

With the rapid adoption of generative AI, a new wave of threats is emerging across the industry with the aim of manipulating the AI systems themselves. One such emerging attack vector is indirect prompt injections. Unlike direct prompt injections, where an attacker directly inputs malicious commands into a prompt, indirect prompt injections involve hidden malicious instructions within external data sources. These may include emails, documents, or calendar invites that instruct AI to exfiltrate user data or execute other rogue actions. As more governments, businesses, and individuals adopt generative AI to get more done, this subtle yet potentially potent attack becomes increasingly pertinent across the industry, demanding immediate attention and robust security measures.


At Google, our teams have a longstanding precedent of investing in a defense-in-depth strategy, including robust evaluation, threat analysis, AI security best practices, AI red-teaming, adversarial training, and model hardening for generative AI tools. This approach enables safer adoption of Gemini in Google Workspace and the Gemini app (we refer to both in this blog as “Gemini” for simplicity). Below we describe our prompt injection mitigation product strategy based on extensive research, development, and deployment of improved security mitigations.


A layered security approach

Google has taken a layered security approach introducing security measures designed for each stage of the prompt lifecycle. From Gemini 2.5 model hardening, to purpose-built machine learning (ML) models detecting malicious instructions, to system-level safeguards, we are meaningfully elevating the difficulty, expense, and complexity faced by an attacker. This approach compels adversaries to resort to methods that are either more easily identified or demand greater resources. 


Our model training with adversarial data significantly enhanced our defenses against indirect prompt injection attacks in Gemini 2.5 models (technical details). This inherent model resilience is augmented with additional defenses that we built directly into Gemini, including: 


  1. Prompt injection content classifiers

  2. Security thought reinforcement

  3. Markdown sanitization and suspicious URL redaction

  4. User confirmation framework

  5. End-user security mitigation notifications


This layered approach to our security strategy strengthens the overall security framework for Gemini – throughout the prompt lifecycle and across diverse attack techniques.


1. Prompt injection content classifiers


Through collaboration with leading AI security researchers via Google's AI Vulnerability Reward Program (VRP), we've curated one of the world’s most advanced catalogs of generative AI vulnerabilities and adversarial data. Utilizing this resource, we built and are in the process of rolling out proprietary machine learning models that can detect malicious prompts and instructions within various formats, such as emails and files, drawing from real-world examples. Consequently, when users query Workspace data with Gemini, the content classifiers filter out harmful data containing malicious instructions, helping to ensure a secure end-to-end user experience by retaining only safe content. For example, if a user receives an email in Gmail that includes malicious instructions, our content classifiers help to detect and disregard malicious instructions, then generate a safe response for the user. This is in addition to built-in defenses in Gmail that automatically block more than 99.9% of spam, phishing attempts, and malware.


A diagram of Gemini’s actions based on the detection of the malicious instructions by content classifiers.


2. Security thought reinforcement


This technique adds targeted security instructions surrounding the prompt content to remind the large language model (LLM) to perform the user-directed task and ignore any adversarial instructions that could be present in the content. With this approach, we steer the LLM to stay focused on the task and ignore harmful or malicious requests added by a threat actor to execute indirect prompt injection attacks.

A diagram of Gemini’s actions based on additional protection provided by the security thought reinforcement technique. 


3. Markdown sanitization and suspicious URL redaction 


Our markdown sanitizer identifies external image URLs and will not render them, making the “EchoLeak” 0-click image rendering exfiltration vulnerability not applicable to Gemini. From there, a key protection against prompt injection and data exfiltration attacks occurs at the URL level. With external data containing dynamic URLs, users may encounter unknown risks as these URLs may be designed for indirect prompt injections and data exfiltration attacks. Malicious instructions executed on a user's behalf may also generate harmful URLs. With Gemini, our defense system includes suspicious URL detection based on Google Safe Browsing to differentiate between safe and unsafe links, providing a secure experience by helping to prevent URL-based attacks. For example, if a document contains malicious URLs and a user is summarizing the content with Gemini, the suspicious URLs will be redacted in Gemini’s response. 


Gemini in Gmail provides a summary of an email thread. In the summary, there is an unsafe URL. That URL is redacted in the response and is replaced with the text “suspicious link removed”. 


4. User confirmation framework


Gemini also features a contextual user confirmation system. This framework enables Gemini to require user confirmation for certain actions, also known as “Human-In-The-Loop” (HITL), using these responses to bolster security and streamline the user experience. For example, potentially risky operations like deleting a calendar event may trigger an explicit user confirmation request, thereby helping to prevent undetected or immediate execution of the operation.


The Gemini app with instructions to delete all events on Saturday. Gemini responds with the events found on Google Calendar and asks the user to confirm this action.


5. End-user security mitigation notifications


A key aspect to keeping our users safe is sharing details on attacks that we’ve stopped so users can watch out for similar attacks in the future. To that end, when security issues are mitigated with our built-in defenses, end users are provided with contextual information allowing them to learn more via dedicated help center articles. For example, if Gemini summarizes a file containing malicious instructions and one of Google’s prompt injection defenses mitigates the situation, a security notification with a “Learn more” link will be displayed for the user. Users are encouraged to become more familiar with our prompt injection defenses by reading the Help Center article. 


Gemini in Docs with instructions to provide a summary of a file. Suspicious content was detected and a response was not provided. There is a yellow security notification banner for the user and a statement that Gemini’s response has been removed, with a “Learn more” link to a relevant Help Center article.

Moving forward


Our comprehensive prompt injection security strategy strengthens the overall security framework for Gemini. Beyond the techniques described above, it also involves rigorous testing through manual and automated red teams, generative AI security BugSWAT events, strong security standards like our Secure AI Framework (SAIF), and partnerships with both external researchers via the Google AI Vulnerability Reward Program (VRP) and industry peers via the Coalition for Secure AI (CoSAI). Our commitment to trust includes collaboration with the security community to responsibly disclose AI security vulnerabilities, share our latest threat intelligence on ways we see bad actors trying to leverage AI, and offering insights into our work to build stronger prompt injection defenses. 


Working closely with industry partners is crucial to building stronger protections for all of our users. To that end, we’re fortunate to have strong collaborative partnerships with numerous researchers, such as Ben Nassi (Confidentiality), Stav Cohen (Technion), and Or Yair (SafeBreach), as well as other AI Security researchers participating in our BugSWAT events and AI VRP program. We appreciate the work of these researchers and others in the community to help us red team and refine our defenses.


We continue working to make upcoming Gemini models inherently more resilient and add additional prompt injection defenses directly into Gemini later this year. To learn more about Google’s progress and research on generative AI threat actors, attack techniques, and vulnerabilities, take a look at the following resources:


Mitigating prompt injection attacks with a layered defense strategy

Posted by Google GenAI Security Team

With the rapid adoption of generative AI, a new wave of threats is emerging across the industry with the aim of manipulating the AI systems themselves. One such emerging attack vector is indirect prompt injections. Unlike direct prompt injections, where an attacker directly inputs malicious commands into a prompt, indirect prompt injections involve hidden malicious instructions within external data sources. These may include emails, documents, or calendar invites that instruct AI to exfiltrate user data or execute other rogue actions. As more governments, businesses, and individuals adopt generative AI to get more done, this subtle yet potentially potent attack becomes increasingly pertinent across the industry, demanding immediate attention and robust security measures.


At Google, our teams have a longstanding precedent of investing in a defense-in-depth strategy, including robust evaluation, threat analysis, AI security best practices, AI red-teaming, adversarial training, and model hardening for generative AI tools. This approach enables safer adoption of Gemini in Google Workspace and the Gemini app (we refer to both in this blog as “Gemini” for simplicity). Below we describe our prompt injection mitigation product strategy based on extensive research, development, and deployment of improved security mitigations.


A layered security approach

Google has taken a layered security approach introducing security measures designed for each stage of the prompt lifecycle. From Gemini 2.5 model hardening, to purpose-built machine learning (ML) models detecting malicious instructions, to system-level safeguards, we are meaningfully elevating the difficulty, expense, and complexity faced by an attacker. This approach compels adversaries to resort to methods that are either more easily identified or demand greater resources. 


Our model training with adversarial data significantly enhanced our defenses against indirect prompt injection attacks in Gemini 2.5 models (technical details). This inherent model resilience is augmented with additional defenses that we built directly into Gemini, including: 


  1. Prompt injection content classifiers

  2. Security thought reinforcement

  3. Markdown sanitization and suspicious URL redaction

  4. User confirmation framework

  5. End-user security mitigation notifications


This layered approach to our security strategy strengthens the overall security framework for Gemini – throughout the prompt lifecycle and across diverse attack techniques.


1. Prompt injection content classifiers


Through collaboration with leading AI security researchers via Google's AI Vulnerability Reward Program (VRP), we've curated one of the world’s most advanced catalogs of generative AI vulnerabilities and adversarial data. Utilizing this resource, we built and are in the process of rolling out proprietary machine learning models that can detect malicious prompts and instructions within various formats, such as emails and files, drawing from real-world examples. Consequently, when users query Workspace data with Gemini, the content classifiers filter out harmful data containing malicious instructions, helping to ensure a secure end-to-end user experience by retaining only safe content. For example, if a user receives an email in Gmail that includes malicious instructions, our content classifiers help to detect and disregard malicious instructions, then generate a safe response for the user. This is in addition to built-in defenses in Gmail that automatically block more than 99.9% of spam, phishing attempts, and malware.


A diagram of Gemini’s actions based on the detection of the malicious instructions by content classifiers.


2. Security thought reinforcement


This technique adds targeted security instructions surrounding the prompt content to remind the large language model (LLM) to perform the user-directed task and ignore any adversarial instructions that could be present in the content. With this approach, we steer the LLM to stay focused on the task and ignore harmful or malicious requests added by a threat actor to execute indirect prompt injection attacks.

A diagram of Gemini’s actions based on additional protection provided by the security thought reinforcement technique. 


3. Markdown sanitization and suspicious URL redaction 


Our markdown sanitizer identifies external image URLs and will not render them, making the “EchoLeak” 0-click image rendering exfiltration vulnerability not applicable to Gemini. From there, a key protection against prompt injection and data exfiltration attacks occurs at the URL level. With external data containing dynamic URLs, users may encounter unknown risks as these URLs may be designed for indirect prompt injections and data exfiltration attacks. Malicious instructions executed on a user's behalf may also generate harmful URLs. With Gemini, our defense system includes suspicious URL detection based on Google Safe Browsing to differentiate between safe and unsafe links, providing a secure experience by helping to prevent URL-based attacks. For example, if a document contains malicious URLs and a user is summarizing the content with Gemini, the suspicious URLs will be redacted in Gemini’s response. 


Gemini in Gmail provides a summary of an email thread. In the summary, there is an unsafe URL. That URL is redacted in the response and is replaced with the text “suspicious link removed”. 


4. User confirmation framework


Gemini also features a contextual user confirmation system. This framework enables Gemini to require user confirmation for certain actions, also known as “Human-In-The-Loop” (HITL), using these responses to bolster security and streamline the user experience. For example, potentially risky operations like deleting a calendar event may trigger an explicit user confirmation request, thereby helping to prevent undetected or immediate execution of the operation.


The Gemini app with instructions to delete all events on Saturday. Gemini responds with the events found on Google Calendar and asks the user to confirm this action.


5. End-user security mitigation notifications


A key aspect to keeping our users safe is sharing details on attacks that we’ve stopped so users can watch out for similar attacks in the future. To that end, when security issues are mitigated with our built-in defenses, end users are provided with contextual information allowing them to learn more via dedicated help center articles. For example, if Gemini summarizes a file containing malicious instructions and one of Google’s prompt injection defenses mitigates the situation, a security notification with a “Learn more” link will be displayed for the user. Users are encouraged to become more familiar with our prompt injection defenses by reading the Help Center article. 


Gemini in Docs with instructions to provide a summary of a file. Suspicious content was detected and a response was not provided. There is a yellow security notification banner for the user and a statement that Gemini’s response has been removed, with a “Learn more” link to a relevant Help Center article.

Moving forward


Our comprehensive prompt injection security strategy strengthens the overall security framework for Gemini. Beyond the techniques described above, it also involves rigorous testing through manual and automated red teams, generative AI security BugSWAT events, strong security standards like our Secure AI Framework (SAIF), and partnerships with both external researchers via the Google AI Vulnerability Reward Program (VRP) and industry peers via the Coalition for Secure AI (CoSAI). Our commitment to trust includes collaboration with the security community to responsibly disclose AI security vulnerabilities, share our latest threat intelligence on ways we see bad actors trying to leverage AI, and offering insights into our work to build stronger prompt injection defenses. 


Working closely with industry partners is crucial to building stronger protections for all of our users. To that end, we’re fortunate to have strong collaborative partnerships with numerous researchers, such as Ben Nassi (Confidentiality), Stav Cohen (Technion), and Or Yair (SafeBreach), as well as other AI Security researchers participating in our BugSWAT events and AI VRP program. We appreciate the work of these researchers and others in the community to help us red team and refine our defenses.


We continue working to make upcoming Gemini models inherently more resilient and add additional prompt injection defenses directly into Gemini later this year. To learn more about Google’s progress and research on generative AI threat actors, attack techniques, and vulnerabilities, take a look at the following resources:


Mitigating prompt injection attacks with a layered defense strategy

Posted by Google GenAI Security Team

With the rapid adoption of generative AI, a new wave of threats is emerging across the industry with the aim of manipulating the AI systems themselves. One such emerging attack vector is indirect prompt injections. Unlike direct prompt injections, where an attacker directly inputs malicious commands into a prompt, indirect prompt injections involve hidden malicious instructions within external data sources. These may include emails, documents, or calendar invites that instruct AI to exfiltrate user data or execute other rogue actions. As more governments, businesses, and individuals adopt generative AI to get more done, this subtle yet potentially potent attack becomes increasingly pertinent across the industry, demanding immediate attention and robust security measures.


At Google, our teams have a longstanding precedent of investing in a defense-in-depth strategy, including robust evaluation, threat analysis, AI security best practices, AI red-teaming, adversarial training, and model hardening for generative AI tools. This approach enables safer adoption of Gemini in Google Workspace and the Gemini app (we refer to both in this blog as “Gemini” for simplicity). Below we describe our prompt injection mitigation product strategy based on extensive research, development, and deployment of improved security mitigations.


A layered security approach

Google has taken a layered security approach introducing security measures designed for each stage of the prompt lifecycle. From Gemini 2.5 model hardening, to purpose-built machine learning (ML) models detecting malicious instructions, to system-level safeguards, we are meaningfully elevating the difficulty, expense, and complexity faced by an attacker. This approach compels adversaries to resort to methods that are either more easily identified or demand greater resources. 


Our model training with adversarial data significantly enhanced our defenses against indirect prompt injection attacks in Gemini 2.5 models (technical details). This inherent model resilience is augmented with additional defenses that we built directly into Gemini, including: 


  1. Prompt injection content classifiers

  2. Security thought reinforcement

  3. Markdown sanitization and suspicious URL redaction

  4. User confirmation framework

  5. End-user security mitigation notifications


This layered approach to our security strategy strengthens the overall security framework for Gemini – throughout the prompt lifecycle and across diverse attack techniques.


1. Prompt injection content classifiers


Through collaboration with leading AI security researchers via Google's AI Vulnerability Reward Program (VRP), we've curated one of the world’s most advanced catalogs of generative AI vulnerabilities and adversarial data. Utilizing this resource, we built and are in the process of rolling out proprietary machine learning models that can detect malicious prompts and instructions within various formats, such as emails and files, drawing from real-world examples. Consequently, when users query Workspace data with Gemini, the content classifiers filter out harmful data containing malicious instructions, helping to ensure a secure end-to-end user experience by retaining only safe content. For example, if a user receives an email in Gmail that includes malicious instructions, our content classifiers help to detect and disregard malicious instructions, then generate a safe response for the user. This is in addition to built-in defenses in Gmail that automatically block more than 99.9% of spam, phishing attempts, and malware.


A diagram of Gemini’s actions based on the detection of the malicious instructions by content classifiers.


2. Security thought reinforcement


This technique adds targeted security instructions surrounding the prompt content to remind the large language model (LLM) to perform the user-directed task and ignore any adversarial instructions that could be present in the content. With this approach, we steer the LLM to stay focused on the task and ignore harmful or malicious requests added by a threat actor to execute indirect prompt injection attacks.

A diagram of Gemini’s actions based on additional protection provided by the security thought reinforcement technique. 


3. Markdown sanitization and suspicious URL redaction 


Our markdown sanitizer identifies external image URLs and will not render them, making the “EchoLeak” 0-click image rendering exfiltration vulnerability not applicable to Gemini. From there, a key protection against prompt injection and data exfiltration attacks occurs at the URL level. With external data containing dynamic URLs, users may encounter unknown risks as these URLs may be designed for indirect prompt injections and data exfiltration attacks. Malicious instructions executed on a user's behalf may also generate harmful URLs. With Gemini, our defense system includes suspicious URL detection based on Google Safe Browsing to differentiate between safe and unsafe links, providing a secure experience by helping to prevent URL-based attacks. For example, if a document contains malicious URLs and a user is summarizing the content with Gemini, the suspicious URLs will be redacted in Gemini’s response. 


Gemini in Gmail provides a summary of an email thread. In the summary, there is an unsafe URL. That URL is redacted in the response and is replaced with the text “suspicious link removed”. 


4. User confirmation framework


Gemini also features a contextual user confirmation system. This framework enables Gemini to require user confirmation for certain actions, also known as “Human-In-The-Loop” (HITL), using these responses to bolster security and streamline the user experience. For example, potentially risky operations like deleting a calendar event may trigger an explicit user confirmation request, thereby helping to prevent undetected or immediate execution of the operation.


The Gemini app with instructions to delete all events on Saturday. Gemini responds with the events found on Google Calendar and asks the user to confirm this action.


5. End-user security mitigation notifications


A key aspect to keeping our users safe is sharing details on attacks that we’ve stopped so users can watch out for similar attacks in the future. To that end, when security issues are mitigated with our built-in defenses, end users are provided with contextual information allowing them to learn more via dedicated help center articles. For example, if Gemini summarizes a file containing malicious instructions and one of Google’s prompt injection defenses mitigates the situation, a security notification with a “Learn more” link will be displayed for the user. Users are encouraged to become more familiar with our prompt injection defenses by reading the Help Center article. 


Gemini in Docs with instructions to provide a summary of a file. Suspicious content was detected and a response was not provided. There is a yellow security notification banner for the user and a statement that Gemini’s response has been removed, with a “Learn more” link to a relevant Help Center article.

Moving forward


Our comprehensive prompt injection security strategy strengthens the overall security framework for Gemini. Beyond the techniques described above, it also involves rigorous testing through manual and automated red teams, generative AI security BugSWAT events, strong security standards like our Secure AI Framework (SAIF), and partnerships with both external researchers via the Google AI Vulnerability Reward Program (VRP) and industry peers via the Coalition for Secure AI (CoSAI). Our commitment to trust includes collaboration with the security community to responsibly disclose AI security vulnerabilities, share our latest threat intelligence on ways we see bad actors trying to leverage AI, and offering insights into our work to build stronger prompt injection defenses. 


Working closely with industry partners is crucial to building stronger protections for all of our users. To that end, we’re fortunate to have strong collaborative partnerships with numerous researchers, such as Ben Nassi (Confidentiality), Stav Cohen (Technion), and Or Yair (SafeBreach), as well as other AI Security researchers participating in our BugSWAT events and AI VRP program. We appreciate the work of these researchers and others in the community to help us red team and refine our defenses.


We continue working to make upcoming Gemini models inherently more resilient and add additional prompt injection defenses directly into Gemini later this year. To learn more about Google’s progress and research on generative AI threat actors, attack techniques, and vulnerabilities, take a look at the following resources: