Google Uses AI Agents to Strengthen Chrome Security, Discovers 13-Year-Old Sandbox Escape
Key Takeaways
- ▸Google's AI vulnerability discovery agent found a critical 13-year-old sandbox escape in Chrome, demonstrating the capability of large language models to detect long-missed security flaws
- ▸The system supports model interoperability, leveraging both open-weight and proprietary models, and includes a 'critic' agent to improve reasoning about security boundaries
- ▸AI vulnerability detection is integrated into Chrome's security infrastructure alongside traditional fuzzing, which remains effective at finding complex, multi-component bugs
Summary
Google announced a major advancement in Chrome security through AI-powered vulnerability discovery. The company has deployed a specialized AI agent harness, powered by Gemini and supporting multiple models, to automatically find security bugs in the Chrome codebase. The system successfully identified a critical sandbox escape vulnerability that had gone undetected for over 13 years—a bug that could allow compromised renderers to access local files. This represents a significant evolution of Google's multi-year AI security research program, which previously collaborated with Project Zero and DeepMind on initiatives like Naptime and Big Sleep. To ensure safety, Google has implemented strict guardrails, running AI models exclusively on isolated, air-gapped machines with restricted network access and preventing model interference with the codebase.
- Strict safety measures limit AI model execution to isolated, air-gapped machines with blocked network access and prevent unauthorized file or system modifications
Editorial Opinion
AI-powered vulnerability discovery represents a paradigm shift for software security, and Google's discovery of a 13-year-old sandbox escape is compelling proof of concept. However, the industry-wide implications are sobering: if LLMs can find years-old vulnerabilities in Chrome's codebase, they can likely do so across millions of open-source and proprietary codebases—raising urgent questions about vulnerability disclosure timelines and how enterprises can scale security testing. Google's transparent approach and emphasis on safety guardrails sets a responsible precedent, but this capability will inevitably diffuse to actors with fewer ethical constraints.
