Our aim is to propel every stage of our process forward as swiftly as possible.
**Identifying Vulnerabilities**
The Chrome Security team has effectively leveraged large language models (LLMs) for several years. In 2023, we pioneered techniques to enhance the coverage and efficiency of security fuzzing through LLMs. The following year, we partnered with Project Zero on a project named Naptime, which equipped LLMs with specialized tools for vulnerability research. In 2025, a collaboration with DeepMind and Project Zero culminated in Big Sleep, an AI entity dedicated to vulnerability discovery that was successful in identifying bugs within the V8 JavaScript engine and the graphics stack.
By early 2026, we developed an advanced agent harness utilizing Gemini, allowing for more efficient vulnerability detection across the Chrome codebase while minimizing false positives. One significant bug we discovered was a sandbox escape issue that could enable a compromised renderer to deceive the browser into accessing local files; astonishingly, this flaw had existed in our codebase undetected for over 13 years. This experience underscored the immense potential of AI-driven vulnerability detection.
We subsequently refined our vulnerability identification harness by:
- Introducing model interoperability to exploit the distinctive strengths of both open-source and proprietary models. - Creating a comprehensive knowledge base of Chrome that includes past CVEs and the complete Git history, enhancing LLMs' reasoning capabilities beyond their initial training data. - Encouraging developers to include SECURITY.md files, which provide models with insights into trust boundaries and the threat landscape. - Implementing a critic agent, designed to review these SECURITY.md files from a different context. - Allowing repeated vulnerability detection runs over the codebase to consider model improvements and non-determinism over time.
All our advancements have been developed with safety as a priority; we have established guardrails to reduce the likelihood of unforeseen AI behavior. Our AI evaluates source code exclusively in a secure environment, utilizing isolated machines devoid of general internet access. Additionally, we deploy a dedicated infrastructure for these internal scans that monitors all network requests with stringent allowlist practices, preventing any suspicious model actions. We also ensure that models do not operate in unrestricted modes, and that our subagents are restricted from altering the local system or accessing files outside designated source code areas.
AI-driven vulnerability detection serves as a valuable enhancement to our existing security testing framework. For instance, fuzzing remains particularly effective at detecting bugs that stem from complex interactions within our codebase or those that arise from seemingly unrelated processes.
We are also committed to recognizing external researchers for their exceptional skills in identifying significant vulnerabilities through the Chrome Vulnerability Reward Program (VRP). By early 2026, we noted a steady increase in bug reports across all categories, with a noticeable leap in March when submissions surpassed the total from all of 2025. This prompted us to refine our VRP to encourage researchers to submit bugs that provide additional insights beyond our internal findings, facilitating seamless integration with our automated processing pipelines.
**Managing Vulnerabilities**
As we utilize AI-enhanced tools to discover an increasing number of security vulnerabilities, we have simultaneously employed AI to automate the processes of validating, prioritizing, and rectifying these bugs. Traditionally, evaluating a single security report could take between 5 minutes to over half an hour, predominantly relying on human expertise. We are progressively automating this triage process, combining rule-based systems with AI to enhance both efficiency and accuracy.
Our automated triage workflow consists of four primary phases:
1. **Noise reduction**: The system filters out spam, confirms whether incoming bugs align with our intake criteria (e.g., no duplicates), and verifies that they delineate a valid Chrome security vulnerability. 2. **Bug replication**: It checks for proof of concept, testing reproducible bugs against the specific operating system and browser versions. This step helps to gather additional details like stack traces that are crucial for remedying the bug. 3. **Enrichment of reports**: The system adds important metadata, including the introduction date of the bug and its severity. To streamline this process, we've clarified our severity guidelines for automatic application. Developers can still adjust severity ratings if needed and provide context via SECURITY.md files for better model understanding. 4. **Automatic assignment**: The system ensures the bug is directed to the appropriate component and human owner.
Although precise metrics are challenging to ascertain, we estimate that this refined process saves hundreds of hours of developer time monthly, enabling our team to concentrate on other security initiatives.
**Resolving Vulnerabilities**
At Google, the responsibility for prioritizing security fixes is shared between developers and the security team, but an effective bug discovery method necessitates an equally robust bug-fixing framework.
To facilitate this, we employ multi-agent workflows:
Following preliminary build steps that provide context for specific issues, a fixing agent generates several candidate solutions. A critic agent then assesses these options, presenting relevant artifacts for developers to review. This iterative process mimics a conventional code review, ensuring compliance with Chromium and Google style guidelines as well as other local code standards.
Agents dedicated to writing tests assist in ensuring that these tests are validated across all supported platforms and configurations of Chrome before being reviewed by developers, saving weeks of potential development time.
Currently, we have LLMs generating candidate fixes for the majority of vulnerabilities, significantly accelerating the pace of security updates in recent Chrome releases.
