Google Open-Sources Mantis, an AI Agent Framework to Reduce Vulnerability Scan False Positives

According to Google, typical AI code scanning suffers from hallucinated bugs and weak true-positive rates below 7%. Mantis aims to improve on this by combining industry-standard agentic techniques, such as critic and review agents, with sandboxed reproduction of vulnerabilities to ground findings in real evidence.

The framework does not blindly scan files. Instead, it analyzes repository history, previous security fixes, architecture, and threat models. It summarizes files into a hierarchical tree containing directory- and repository-level context, which Google says reduces token usage by 85% while retaining important structural information.

Mantis uses a strategist agent to evaluate high-level code structure, threat models, and dependency graphs, alongside research agents that examine raw source files in depth, tracing data flows, control flows, and sanitization logic. The system is organized as a modular skill suite with over 15 tools that can be executed sequentially or in parallel. Stages communicate by reading and writing to a shared state stored on disk.

The framework supports multiple AI models and allows combining different models for different phases of the scanning pipeline. Mantis is part of Google's internal effort to find and fix vulnerabilities at machine speed.

Why This Matters

Reduces developer noise: Security teams are overwhelmed by false positives from automated scanners; Mantis aims to surface only real vulnerabilities, saving time and effort.
Brings grounding to AI security tools: By requiring sandboxed reproduction, Mantis moves beyond pure LLM judgement, potentially setting a new standard for trustworthy AI-assisted vulnerability detection.
Open-source availability: The release allows the wider security community to inspect, adapt, and improve the framework, accelerating progress in automated secure development.

Background

AI-powered code scanning tools have grown popular but often suffer from high false-positive rates and confidence in hallucinated vulnerabilities. Google's own research indicated true-positive rates under 7% for conventional approaches. The company has been investing in agentic AI systems that combine multiple specialised agents (critic, reviewer, strategist) to improve reliability. Mantis is the first such framework from Google to be released as open source, drawing on internal experience from its security teams.

Key Perspectives

Security developers and DevOps teams: They stand to benefit from reduced noise and faster triage, but must integrate Mantis into existing pipelines and evaluate its performance against other tools. Google: By open-sourcing Mantis, Google positions itself as a leader in practical, grounded AI for security while also benefiting from community contributions and external validation. Critics and skeptics: Some may question whether Mantis's reliance on LLMs—even with sandboxed grounding—can fully eliminate hallucinations or whether the framework creates a false sense of security. The need for careful model selection and customisation may also limit adoption for smaller teams.

What to Watch

Adoption metrics: How many organisations integrate Mantis into their CI/CD pipelines and how its true-positive rate compares in real-world deployments.
Community contributions: Whether external researchers find and patch limitations, improving Mantis's robustness over time.
Competing frameworks: Watch for similar agentic approaches from other cloud providers or security vendors, and whether Mantis becomes a de facto standard or remains one option among many.

ZOTPAPER