Google Cloud’s AI Agents Now Patch Their Own Security Bugs

Google Cloud’s security team has unveiled a multi-agent AI framework that autonomously finds, validates, and patches vulnerabilities across its software development lifecycle. The system, detailed by CISO Chris Betz and senior director Ruchi Shah, aims to counter AI-powered threats at machine speed.

From Manual Gates to Agentic Pipelines

Traditional security reviews are slow and manual. Google Cloud now routes product launches through an agent-based pipeline that cross-references designs against a catalog of over 200 security requirements. High-risk items are automatically flagged for human review, while a dynamic product dossier replaces static threat models.

Mantis: Open-Source Multi-Agent Scanning

Naive AI code scanning often hallucinates bugs, with true-positive rates under 7%. Google’s Mantis framework solves this with a hierarchical security summary tree that reduces token overhead by 85% while preserving context. The core skills are now open source.

Mantis uses a coordinated workflow:

• Strategist agent: Evaluates code structure and threat models to prioritize investigation tasks.
• Research agents: Drill into source files to examine data flows and sanitization logic.
• Deduplicator, reviewer, and critic agents: Filter noise and eliminate false positives.
• Reproduction sandbox: Automatically runs AI-generated proof-of-concept exploits to verify real-world exploitability.

Self-Healing Fuzz Testing

Writing fuzz harnesses is a bottleneck. Google’s autonomous engine eliminates manual intervention with Context and Drafting agents that author harnesses, Building and Testing agents that fix compilation errors via a Hallucination Cleaner, and Quality Analyzer agents that adjust inputs to penetrate deeper APIs.

Unified AI Patching Pipeline

Finding thousands of vulnerabilities creates a remediation backlog. Google’s pipeline routes findings into an autonomous fix cycle: a Reproduce agent replicates the crash, a Bug Context agent maps the failure path, a Patch agent generates a fix, and an Evaluation agent runs regression tests before human review.

Continuous Self-Reflection

Stateless AI systems repeat mistakes. Google’s framework adds a post-hoc self-reflection loop that analyzes execution logs and human feedback, storing successful patterns in a global knowledge store. This compounding intelligence improves fix success rates and efficiency.

Google Cloud is pioneering “immune software” that continuously discovers, validates, and patches its own weaknesses. For more on Mantis and AI threat defense, check out the full post.

If you’re building on Google Cloud, you might also enjoy our Google Cloud review and model pricing comparison.

Original announcement published on Google Cloud.