Google Mantis reduces false positives in scanning

Google has made public the source code for Mantis, an AI‑agent framework that automates the entire vulnerability‑scanning lifecycle, from initial detection through to remediation.

Framework design and agent roles

The architecture relies on a collection of specialized components that exchange information via a shared disk state. A “summarize” module builds a hierarchical view of a repository, preserving directory and architectural context while cutting token usage by roughly 85 %.

Separate “critic” and “review” units evaluate findings, filtering out low‑confidence results. A “strategist” component examines overall code structure, threat models, and dependency graphs, while research agents perform deep searches of raw source files to trace data and control flows.

Cutting down false positives

Conventional AI code scanners often generate hallucinated bugs, with true‑positive rates reported below 7 %. Mantis addresses this problem by grounding each finding in a sandboxed reproduction step, providing concrete evidence that a defect can be triggered safely.

The review stage applies a rule‑based negative filter to weed out likely false alerts. Google warns that the filter should be used carefully; discarding low‑risk findings indiscriminately could hide genuine issues.

Compared with earlier attempts that relied mainly on large language models scanning files line‑by‑line, this layered approach mirrors manual security reviews more closely. By incorporating repository history and prior fixes, the framework gains context that simple pattern matching lacks, which should translate into fewer wasted tickets for developers.

Choosing models for each task

Mantis supports multiple AI models and lets users pair model size with task complexity. For quick classification or clustering, the documentation recommends “flash” or “lite” variants, which run faster and cost less.

Tasks that need deeper reasoning—such as generating a reproducible crash or crafting a side‑effect‑free patch, use more capable models. The system can even combine different models across stages, avoiding the need to run the most powerful model on every step.

The open‑source release includes over 15 modular tools, each callable individually or as part of a pipeline. Detailed contracts between stages and best‑practice guidance are available in the agent reference guide.

Availability and next steps

The code is hosted on GitHub, where developers can clone the repository and explore the full suite of agents. Documentation links to Google’s internal approach to rapid vulnerability detection for additional context.

Early adopters will need to configure model endpoints, set up sandbox environments, and tune the negative filter to balance coverage with noise reduction. As the community builds on the framework, real‑world feedback will shape future refinements.

Developers can now test fixes quickly.