Extended Stable Updates for Desktop

The Extended Stable channel has been updated to 146.0.7680.216 for Windows and Mac which will roll out over the coming days/weeks.


A full list of changes in this build is available in the log. Interested in switching release channels? Find out how here. If you find a new issue, please let us know by filing a bug. The community help forum is also a great place to reach out for help or learn about common issues.

Srinivas Sista
Google Chrome

Stable Channel Update for Desktop

The Stable channel has been updated to 147.0.7727.137/138 for Windows/Mac  and 147.0.7727.137 for Linux, which will roll out over the coming days/weeks. A full list of changes in this build is available in the Log

Security Fixes and Rewards

Note: Access to bug details and links may be kept restricted until a majority of users are updated with a fix. We will also retain restrictions if the bug exists in a third party library that other projects similarly depend on, but haven’t yet fixed.


This update includes 30 security fixes. Below, we highlight fixes that were contributed by external researchers. Please see the Chrome Security Page for more information.


[$7000][494352590] Critical CVE-2026-7363: Use after free in Canvas. Reported by heapracer on 2026-03-19

[TBD][493221953] Critical CVE-2026-7361: Use after free in iOS. Reported by Google on 2026-03-16

[TBD][503419515] Critical CVE-2026-7344: Use after free in Accessibility. Reported by Google on 2026-04-16

[TBD][503645680] Critical CVE-2026-7343: Use after free in Views. Reported by Google on 2026-04-17

[$16000][493955227] High CVE-2026-7333: Use after free in GPU. Reported by c6eed09fc8b174b0f3eebedcceb1e792 on 2026-03-19

[TBD][495852034] High CVE-2026-7360: Insufficient validation of untrusted input in Compositing. Reported by Google on 2026-03-24

[TBD][496284494] High CVE-2026-7359: Use after free in ANGLE. Reported by Google on 2026-03-25

[TBD][496285281] High CVE-2026-7358: Use after free in Animation. Reported by Google on 2026-03-25

[TBD][496456528] High CVE-2026-7334: Use after free in Views. Reported by Batuhan Eşref KOÇ on 2026-03-26

[TBD][497047552] High CVE-2026-7357: Use after free in GPU. Reported by Google on 2026-03-27

[TBD][497769116] High CVE-2026-7356: Use after free in Navigation. Reported by Google on 2026-03-30

[TBD][498746519] High CVE-2026-7354: Out of bounds read and write in Angle. Reported by Google on 2026-04-01

[TBD][498809718] High CVE-2026-7353: Heap buffer overflow in Skia. Reported by Google on 2026-04-01

[TBD][499023054] High CVE-2026-7352: Use after free in Media. Reported by Google on 2026-04-02

[TBD][499119490] High CVE-2026-7351: Race in MHTML. Reported by Google on 2026-04-02

[TBD][500018484] High CVE-2026-7350: Use after free in WebMIDI. Reported by Google on 2026-04-06

[TBD][500034684] High CVE-2026-7349: Use after free in Cast. Reported by Google on 2026-04-06

[TBD][500104917] High CVE-2026-7348: Use after free in Codecs. Reported by Google on 2026-04-06

[TBD][500387779] High CVE-2026-7335: Use after free in media. Reported by Jungwoo Lee (@physicube) and Wongi Lee (@_qwerty_po) on 2026-04-07

[TBD][500767595] High CVE-2026-7336: Use after free in WebRTC. Reported by Mozilla on 2026-04-09

[TBD][500880819] High CVE-2026-7337: Type Confusion in V8. Reported by [email protected] on 2026-04-09

[TBD][501722605] High CVE-2026-7347: Use after free in Chromoting. Reported by Google on 2026-04-11

[TBD][502206907] High CVE-2026-7346: Inappropriate implementation in Tint. Reported by Google on 2026-04-13

[TBD][502248774] High CVE-2026-7345: Insufficient validation of untrusted input in Feedback. Reported by Google on 2026-04-13

[TBD][502449857] High CVE-2026-7338: Use after free in Cast. Reported by Krace on 2026-04-14

[TBD][503889643] High CVE-2026-7342: Use after free in WebView. Reported by Google on 2026-04-17

[TBD][504586599] High CVE-2026-7341: Use after free in WebRTC. Reported by Google on 2026-04-20

[$4000][493957495] Medium CVE-2026-7339: Heap buffer overflow in WebRTC. Reported by c6eed09fc8b174b0f3eebedcceb1e792 on 2026-03-19

[$3000][497896137] Medium CVE-2026-7340: Integer overflow in ANGLE. Reported by 86ac1f1587b71893ed2ad792cd7dde32 on 2026-03-30

[TBD][498285711] Medium CVE-2026-7355: Use after free in Media. Reported by Google on 2026-03-31


We would also like to thank all security researchers that worked with us during the development cycle to prevent security bugs from ever reaching the stable channel.

Many of our security bugs are detected using AddressSanitizer, MemorySanitizer, UndefinedBehaviorSanitizer, Control Flow Integrity, libFuzzer, or AFL.


Interested in switching release channels? Find out how here. If you find a new issue, please let us know by filing a bug. The community help forum is also a great place to reach out for help or learn about common issues.

Srinivas Sista

Google Chrome

Students can now create personal class notebooks with NotebookLM in Google Classroom

Last year, we announced that educators can use NotebookLM to create teacher-led AI experiences for students in Google Classroom, based on their class materials. Since then, students have been able to access their teachers’ interactive study guides and other learning aids created with NotebookLM for extra practice, support, and learning opportunities. Today, we’re allowing students in higher education who are 18 years of age and older to create their own notebooks for courses in Google Classroom.

From the Gemini tab in Google Classroom, students can use NotebookLM to create a personal class notebook that is grounded in the materials provided by their educator. Personal class notebooks can transform how students interact with these class resources, unlocking new ways to build understanding by converting class materials into interactive and multi-modal study tools. This can help students with:

  • Creating custom study tools: Students can use the "Studio" panel within their notebook to generate various high-value outputs, including Audio Overviews (podcast-style summaries), Video Overviews, study guides, flashcards, and interactive visual diagrams.
  • Summarizing and synthesizing: Students can quickly synthesize information across up to 50 source documents per notebook, making it easier to prepare for exams or catch up on missed lessons.
  • Direct-to-student support: By using the Gemini tab directly in Classroom, students can ask questions and get grounded answers based strictly on their class content, ensuring help is relevant and reliable.
  • Enhanced creativity: Students can go beyond text by creating infographics, slide decks, and other visual aids to help them internalize and present what they’ve learned.
This feature will roll out on the web first, with mobile following in the coming weeks.

Getting started

Rollout pace

Availability

  • Education: Education Fundamentals, Standard, and Plus

Resources

Students can now create personal class notebooks with NotebookLM in Google Classroom

Last year, we announced that educators can use NotebookLM to create teacher-led AI experiences for students in Google Classroom, based on their class materials. Since then, students have been able to access their teachers’ interactive study guides and other learning aids created with NotebookLM for extra practice, support, and learning opportunities. Today, we’re allowing students in higher education who are 18 years of age and older to create their own notebooks for courses in Google Classroom.

From the Gemini tab in Google Classroom, students can use NotebookLM to create a personal class notebook that is grounded in the materials provided by their educator. Personal class notebooks can transform how students interact with these class resources, unlocking new ways to build understanding by converting class materials into interactive and multi-modal study tools. This can help students with:

  • Creating custom study tools: Students can use the "Studio" panel within their notebook to generate various high-value outputs, including Audio Overviews (podcast-style summaries), Video Overviews, study guides, flashcards, and interactive visual diagrams.
  • Summarizing and synthesizing: Students can quickly synthesize information across up to 50 source documents per notebook, making it easier to prepare for exams or catch up on missed lessons.
  • Direct-to-student support: By using the Gemini tab directly in Classroom, students can ask questions and get grounded answers based strictly on their class content, ensuring help is relevant and reliable.
  • Enhanced creativity: Students can go beyond text by creating infographics, slide decks, and other visual aids to help them internalize and present what they’ve learned.
This feature will roll out on the web first, with mobile following in the coming weeks.

Getting started

Rollout pace

Availability

  • Education: Education Fundamentals, Standard, and Plus

Resources

Introducing AMS: Activation-based model scanner for open-weight LLM safety verification

by Glen Messenger, Google Kubernetes Engine (GKE)

The open-weight model ecosystem is thriving—and so is its shadow. A 2025 study identified over 8,000 safety-modified model repositories on Hugging Face alone, with modified models complying with unsafe requests at rates of 74% compared to 19% for their original instruction-tuned counterparts.

For organizations deploying open-weight models, a critical question emerges: how do you know the model you downloaded is safe to run?

We believe defensive security tools should be widely available. AMS represents our contribution to a safer AI ecosystem—one where developers everywhere can verify model integrity before deployment.

Today we're releasing AMS (Activation-based Model Scanner), an open source tool that answers this question in 10–40 seconds—without sending a single prompt.

The Problem with Behavioral Testing

Traditional safety verification relies on behavioral testing: send harmful prompts, check if the model refuses. This approach has three fundamental limitations.

It's slow. Comprehensive benchmarks like HarmBench require hundreds of queries. For organizations running continuous integration pipelines or screening large model registries, this can be impractical.

It's incomplete. No benchmark covers every harmful behavior. Models can exhibit safe behavior on known test sets while remaining unsafe on novel or out-of-distribution prompts.

It's gameable. Models can be fine-tuned to refuse benchmark prompts while complying with novel attacks—a known limitation of purely behavioral evaluation approaches.

A Structural Approach

AMS scanner validating clean and tampered models at select layers of the model stack, using activation geometry comparisons to detect anomalies
Clean vs Tampered Models

AMS takes a different approach entirely. Instead of testing what a model says, it measures how a model thinks.

Safety training creates measurable geometric structure in a model's activation space. Instruction-tuned models develop internal "direction vectors"—representations that separate harmful content from benign content with high statistical confidence (4–8σ separation). When safety training is removed—through fine-tuning, abliteration, or training on unfiltered data—this geometric structure collapses.

AMS measures this collapse directly. The approach is grounded in recent research on representation engineering, which demonstrates that high-level concepts are encoded linearly in LLM activation space and can be reliably extracted via simple linear probes on intermediate-layer hidden states.

git clone https://github.com/GoogleCloudPlatform/activation-model-scanner.git
cd activation-model-scanner && pip install -e .

# Standard scan (3 concepts: harmful_content, injection_resistance, refusal_capability)
ams scan ./my-model

# Quick scan (2 concepts, ~40% faster)
ams scan ./my-model --mode quick

# Full scan (4 concepts including truthfulness)
ams scan ./my-model --mode full

# JSON output for CI/CD pipelines
ams scan ./my-model --json

What AMS Detects

AMS operates as a two-tier scanner. Tier 1 measures whether safety-relevant activation structure exists at all—no baseline required. Tier 2 compares a model's activation fingerprint against a verified baseline to detect subtle modifications, including supply chain substitution.

In our validation across 14 model configurations:

  • Instruction-tuned models (Llama, Gemma, Qwen) show 3.8–8.4σ separation—consistent with strong safety training
  • Uncensored variants (Dolphin, Lexi) show collapsed separation at 1.1–1.3σ—flagged as CRITICAL
  • Abliterated models show partial degradation at 3.3σ—flagged as WARNING
  • Base models (no safety training) show 0.69σ—confirming the absence of safety structure
  • Quantized models (INT4/INT8) show less than 5% separation drift—safe to scan production deployments

Use Cases

Diagram showing three threat vectors : fine-tuned backdoors (hidden trigger behaviours), weight poisoning (direct parameter edit) and supply chain swap (substituted checkpoint)
Threat Landscape

CI/CD Safety Gates

Integrate AMS into your model deployment pipeline to block unsafe models before they reach production. An example Github Actions workflow:

jobs:
model-safety-check:
  runs-on: ubuntu-latest
  steps:
    - uses: actions/checkout@v3

    - name: Install AMS
      run: pip install ams-scanner[cli]

    - name: Scan model
      run: |
        ams scan ./model \
          --verify meta-llama/Llama-3-8B-Instruct \
          --json > scan-results.json

    - name: Upload results
      uses: actions/upload-artifact@v3
      with:
        name: ams-scan-results
        path: scan-results.json

Supply Chain Verification

Confirm that downloaded weights match their claimed identity using Tier 2 fingerprint comparison.

# First, create a baseline from the official model
ams baseline create ./my-model

# Then verify an unknown model against it
ams scan ./suspicious-model --verify ./my-model

Registry Screening

Automatically screen models at upload or download time to flag degraded safety structure before deployment.

# Standard scan (3 concepts: harmful_content, injection_resistance, refusal_capability)
ams scan ./my-model

# Quick scan (2 concepts, ~40% faster)
ams scan ./my-model --mode quick

# Full scan (4 concepts including truthfulness)
ams scan ./my-model --mode full

# JSON output for CI/CD pipelines
ams scan ./my-model --json

How It Works

AMS processes a set of contrastive prompt pairs—examples that differ only in whether they contain harmful content—through the model under inspection. It extracts hidden states at an intermediate layer (typically 35–40% depth), computes a direction vector that separates the two classes, and measures class separation as a σ score.

Flowchart illustrating AMS scanning process: contrastive prompt pairs enter the model, hidden states are extracted at an intermediate layer, direction vectors are computed, and class separation is measured to produce PASS, WARNING, or CRITICAL results
How it Works

The key insight is that this measurement requires no generation, no benchmark queries, and no ground-truth labels. The entire scan completes in a single forward pass per prompt pair, typically 10–40 seconds on GPU hardware.

The probe consists of a single direction vector (~16KB for standard 4096-dimensional models). No model weights are modified. The tool works with any Hugging Face-compatible model.

Get Started

AMS is available now under Apache 2.0:

We welcome contributions, baseline additions for new model families, and feedback from the communities. See the contributing guide in the repository for details.