Google Open-Sources Mantis Agentic Framework to Reduce False Positives in Automated Vulnerability Scanning

Article Summary

Google has open-sourced Mantis, an AI-agent framework designed to automate vulnerability discovery, validation, reproduction, and fixing, aiming to reduce false positives common in AI code scanning. Mantis analyzes repository history, security fixes, and threat models, summarizing them into a hierarchical tree that reduces token usage by 85% while maintaining context. It employs critic and reviewer agents to filter findings and uses sandboxed reproduction for validation, unlike AI scanners that can produce false positives at rates below 7%. Mantis is modular, featuring over 15 tools that can be orchestrated with different AI model classes for specific tasks, optimizing speed and efficiency.

AI Analysis

## Core Assessment
Google’s release of the Mantis framework shifts the burden of vulnerability management from simple pattern matching to a modular, agentic workflow. By utilizing a hierarchical tree structure to compress repository context, the system claims an 85% reduction in token usage—a significant operational optimization for those paying by the token. The framework’s value lies in its "sandboxed reproduction" stage, which moves beyond the probabilistic guessing of standard LLMs to provide actual evidence of a vulnerability. However, the system’s effectiveness hinges on the "mantis-review" stage; as Google notes, over-tuning these rule-based filters risks silencing genuine alerts. It is a classic engineering trade-off: you are trading the simplicity of a standalone scanner for the complexity of maintaining a multi-stage, state-sharing agentic architecture.

## Impact and Advice
* **Security Engineers:** Evaluate the "mantis-reproduce" and "mantis-patch" modules against current manual triage workflows. Do not deploy the negative filter in production until you have calibrated it against historical false-negative rates.
* **Platform Architects:** Review the framework’s disk-based state sharing to ensure it fits within your existing CI/CD ephemeral environment constraints.
* **AI Operations Leads:** Implement the recommended model-tiering strategy—using "flash" models for research and clustering while reserving frontier models for reproduction—to avoid unnecessary compute costs.

## Contrarian View
The modularity of Mantis requires orchestrating over 15 distinct tools that read and write to a shared state on disk. This introduces a non-trivial failure surface: if one agent writes an incorrect state or if inter-stage contracts break due to repository evolution, the entire pipeline could fail silently. Relying on "machine speed" vulnerability fixing is only as good as the underlying threat models and architecture maps provided to the agents; garbage in, agentic garbage out.

## Bottom Line
Mantis provides a structured, multi-model approach to reducing AI-generated vulnerability noise, but it replaces the "false positive" problem with an "orchestration and maintenance" problem. If your team isn't prepared to tune agent-specific filters, you are just trading one set of headaches for another.

Explore this story further