Changing the game: How Google uses agentic AI to secure hundreds of millions of lines of code

AI is accelerating software development, but it also introduces new security risks as AI-generated code becomes more common. Google’s AI and Infrastructure team has developed AI-native agentic methods to embed continuous, high-precision vulnerability scanning and patching directly into its software development lifecycle. By scanning every code change in real time, the approach prevents hundreds of vulnerabilities per month from reaching production, addressing the limitations of traditional one-off security scans that often identify issues too late.

Traditional security scans rely on large, infrequent assessments that lack sufficient context and slow down development. Google’s method shifts to pre-submit scanning, where each code check-in is evaluated in real time using AI agents. This integration into existing developer tools makes security a continuous process, similar to code reviews or style checks. The approach reduces the context needed for scanning, improving effectiveness by focusing on individual changes rather than large, static scans.

To enhance precision, Google evolved Mantis, an open-source multi-agent review harness, by pairing security agents with localized threat models derived from live codebase metadata. These models use dependency call graphs to refine threat contexts, reducing false positives to as low as 3% in some cases. The system encourages developers to keep threat models updated alongside code changes, ensuring ongoing accuracy and relevance.

The solution includes a two-step validation process to maintain low latency: a quick lightweight scan followed by a post-submit scan during nightly integration testing. An automated bug-fix agent then uses scan results to generate precise fixes, which are submitted for human review. This reduces the time between detection and resolution. Google’s approach has significantly improved security, with widespread adoption of its suggestions and prevention of numerous vulnerabilities from entering the codebase.