Google Open-Sources Mantis: Agentic Security Scanner That Cuts False Positives in AI-Powered Vulnerability Detection

Google has released Mantis, an open-source AI-agent framework that automates security vulnerability detection by combining intelligent agent workflows with sandboxed proof-of-concept execution. The system tackles a persistent weakness in conventional AI code scanners: the tendency to flag non-existent bugs and generate hallucinatory vulnerability reports, typically achieving true-positive rates below 7%.

The False-Positive Problem#

Traditional AI-powered security scanners struggle with noise. They analyze code files through brute-force pattern matching and rely heavily on language-model judgments, often producing misleading findings that waste security teams' time. Mantis rethinks this approach by grounding AI assessments in tangible evidence rather than probabilistic guessing.

How Mantis Works#

The framework uses a modular multi-agent architecture that mirrors how human security researchers operate:

• Context-aware analysis: Rather than scanning files indiscriminately, Mantis examines repository history, previous fixes, system architecture, and threat models to focus its search.
• Token-efficient summarization: It builds a hierarchical directory tree representing the codebase structure, reducing token consumption by 85% while preserving essential context.
• Specialized agent roles: A strategist agent evaluates high-level code structure and threat landscapes; researcher agents trace data and control flows through raw source code; critic and reviewer agents filter findings to eliminate false alarms.
• Sandboxed validation: Unlike systems that stop at predictions, Mantis reproduces suspected vulnerabilities in isolated environments to confirm they're exploitable, providing concrete evidence rather than theoretical risk.

Modular Design and Model Selection#

The system comprises over 15 sequential or parallel-executable tools, including mantis-summarize, mantis-review, mantis-critic, and others that share state through disk-based communication. Google recommends pairing lighter model variants (flash or lite) with lightweight tasks like classification and deduplication, while reserving heavier frontier models for complex reasoning—such as mantis-reproduce, which generates working crash proofs, and , which creates safe code fixes.