ReleaseBytes
gcp Google Cloud Blog ·

Google details agentic AI for securing infrastructure code, open-sources Mantis

blogaigcpengineergovernment
announcement

Google's AI and Infrastructure team has developed AI-native agentic methods to embed high-precision, pervasive vulnerability scanning and patching directly into its software development lifecycle. This continuous pre-submit scanning across hundreds of millions of lines of code prevents hundreds of vulnerabilities monthly, safeguarding Google's global network and AI infrastructure. Key innovations include localized threat models, specialized triage agents for low-latency validation, and automated bug-fix agents that propose fixes for human review. To enable other organizations to adopt similar AI-native security transformations, Google has also open-sourced its multi-agent review harness, Mantis.

  • →AI-native agentic methods secure Google's infrastructure code
  • →Pervasive pre-submit scanning with localized threat models and triage agents
  • →Automated bug-fix agents for rapid vulnerability remediation
  • →Mantis multi-agent review harness released as open source
Notes (4) ›
  • AI-native agentic methods secure Google's infrastructure code

    Google has developed and deployed AI-native agentic methods to integrate high-precision vulnerability scanning and patching into its software development lifecycle. This approach continuously scans every code change across Google's massive codebase, preventing hundreds of vulnerabilities monthly from reaching production.

  • Pervasive pre-submit scanning with localized threat models and triage agents

    The core of Google's system is pervasive pre-submit scanning, evaluating each code check-in in real-time using AI agents and localized threat models that leverage live codebase metadata. Specialized triage agents programmatically validate findings, achieving over 92% precision with low latency, supported by a post-submit scan as a second layer of defense.

  • Automated bug-fix agents for rapid vulnerability remediation

    To complete the loop, an automated bug-fix agent constructs precise fixes based on scan results and generated proofs. These fixes are submitted for human review as part of the original change request, significantly reducing the time between vulnerability detection and resolution.

  • Mantis multi-agent review harness released as open source

    Google is making its multi-agent review harness, Mantis, available as open source to help other organizations implement similar AI-native security transformations. Mantis helps compensate for model variability and provides a framework for embedding continuous security scanning into the development lifecycle.

Read the original announcement →

https://cloud.google.com/blog/topics/systems/using-ai-agents-to-secure-google-infrastructure/

Related releases