Google’s Chrome Security Team has detailed a sweeping overhaul of its vulnerability management pipeline, leaning heavily on AI to accelerate bug discovery, triage, and patching as attackers increasingly weaponize LLMs for automated exploit development.
Chrome’s AI security journey started in 2023 with LLM-assisted fuzzing, followed by Project Zero’s Naptime in 2024 and Big Sleep in 2025, which found real bugs in the V8 engine and Chrome’s graphics stack.
In early 2026, Google deployed a Gemini-powered agent harness across the full Chrome codebase, uncovering a sandbox escape (tracked as crbug.com/487383169) that had lurked undetected for over 13 years.
Google Hardens Chrome Against AI-Era Threats
Since then, the team has expanded model interoperability across open weights and proprietary LLMs, built a knowledge base spanning Chrome’s entire Git history and CVE archive, and introduced SECURITY.md files to help models understand trust boundaries.
A dedicated “critic” agent now validates findings, and scans are repeated multiple times to account for model non-determinism. Strict guardrails keep these agents sandboxed, with no internet access, allowlisted network requests, and no ability to modify systems outside designated source directories.
Manual triage once took 5 to 30 minutes per report. Chrome’s new pipeline automates the process across four phases, filtering spam and duplicates, reproducing bugs with stack trace generation, enriching reports with severity metadata, and automatically routing issues to the right developer.
Google estimates this saves hundreds of developer hours every month. On the fix side, multi-agent workflows now generate candidate patches, evaluate them using critic agents, and automatically generate tests across all supported platforms.
The results are dramatic: Chrome 149 and 150 combined fixed 1,072 security bugs, more than the prior 23 milestones combined.
Integration with CodeMender and Big Sleep in continuous integration has already blocked more than 20 vulnerabilities before they reached production, including one critical S1+ issue in May.
To combat “N-day” exploitation, in which attackers reverse-engineer public fixes before patches reach users, Chrome is piloting two security releases per week as part of its broader transition to a two-week major release cadence.
On the client side, Google is developing “dynamic patching” to replace background processes like the Renderer and GPU without requiring a full restart. Chrome 150 already introduced auto-restart on macOS when no windows are open and an update is pending.
Beyond fixing individual bugs, Chrome is pursuing memory safety at scale. MiraclePtr and MiracleObject are being expanded to neutralize up to 90% of use-after-free bugs on the GPU main thread, while a “spanification” effort has brought 97% of first-party code to compile under strict unsafe-buffer warnings.
Longer-term, Google is shifting toward Rust, targeting high-bug-density components like parsers and codecs for rewrite. The company is also extending automated dependency updates across Chrome’s more than 2,300 third-party libraries, using safety signals from its GOSSIP platform to vet upstream risk.
Google frames the rising bug-fix count not as a red flag but as evidence the defensive machinery is working, closing the window attackers have to exploit vulnerabilities before users are protected.
Cut SOC investigation blind spots and contain threats earlier to reduce response costs and business disruption with ANY.RUN.
