Chrome’s AI security loop finds a 13-year-old sandbox escape

Chrome’s security team is moving AI deeper into the browser’s vulnerability workflow. Google says an early-2026 Gemini-based agent harness found issues across the broader Chrome codebase with better efficiency and fewer false positives, including a sandbox escape that had survived for more than 13 years. The bug could have let a compromised renderer trick the browser into reading local files.

The notable part is the operating model. Chrome’s team traces the work back to 2023 experiments using LLMs to improve security fuzzing coverage and performance. In 2024, Project Zero’s Naptime gave LLMs specialized tools for vulnerability research. In 2025, Big Sleep, built with DeepMind and Project Zero, found bugs in V8 and the graphics stack. The new post frames those experiments as part of a larger loop for finding, triaging, fixing, and shipping security updates faster.

Google lists several practical changes: model interoperability across open-weight and proprietary systems, a Chrome knowledge base built from past CVEs and Git history, and better developer workflows around agent-produced findings. For security teams, the value is not just whether a model flags a bug. It is whether the system can produce evidence, keep false positives manageable, and fit into the review path that turns a finding into a shipped patch.

HN discussion pushed on the same issues. Some commenters connected the findings to C and C++ memory-safety risk, while others asked how much of the result came from AI versus a concentrated internal security push. The useful takeaway sits between those views: AI is becoming part of the defensive supply chain, and browsers are large enough that shaving time from each step of a vulnerability’s life cycle matters.