Google Mantis: An Agentic Vulnerability Scanning Harness for Reducing False Positives

Details

Google released Mantis as open source, an agentic framework designed to automate the full vulnerability lifecycle: identification, validation, sandboxed reproduction, and patching. It directly addresses a documented weakness in AI-powered code scanning — conventional tools produce hallucinated bugs and true-positive rates routinely under 7%, generating noise that overwhelms security teams. The framework is organized as a modular pipeline with distinct agent roles. A strategist agent evaluates high-level repository structure, threat models, and dependency graphs. Research agents trace data flows, control flows, and sanitization logic through raw source code. Critic and reviewer agents filter findings before they surface. A hierarchical tree summarization step condenses repository context into structured form, cutting token consumption by 85% while preserving architectural signal. The key differentiator is sandboxed reproduction: Mantis can execute a suspected vulnerability in a controlled environment before flagging it, providing executable evidence rather than relying on LLM judgment alone. The 15+ modular tools — including mantis-summarize, mantis-review, and mantis-critic — communicate via shared disk state and can run sequentially or in parallel. Different model tiers can be assigned to different stages, pairing lightweight models with fast screening tasks and heavier models with deep analysis. Mantis is part of Google's internal program to find and fix vulnerabilities at machine speed. Now open-sourced, enterprise security teams can deploy it without dependency on Google-proprietary services.