Mantis is a modular toolkit of security review skills for AI coding agents to autonomously find, reproduce, and patch vulnerabilities. It is designed to run in isolated environments and supports multiple agent frameworks.
Collecting history — the radar snapshots this repo daily. The trend line appears after 3 days of data (2 so far).
What it is
Mantis Skills: Portable Toolkit for Building Security Review Harnesses. It is a decoupled, sequential, and security-focused set of skills designed for use with Coding Agents to review codebases for vulnerabilities. The project is intended as a starting point to adapt, tune, and extend to fit different stacks and domains.
How it works
The toolkit provides a pipeline of security review steps and knowledge management for autonomous agent workflows. The README references an Agent Reference Guide for detailed breakdown of pipeline stages, inter-stage contracts, and living threat model development. It emphasizes operating in isolated environments and using human-in-the-loop where appropriate. It also notes the architecture supports multiple agent frameworks and can integrate with tools like Gemini CLI and Antigravity CLI, among others.
Getting started
Prerequisites and setup suggestions include configuring a local CLI environment, using containers for testing, and optionally enabling gVisor runtimes in Docker for extra security. The README provides commands to install the skills via CLI:
npx skills add google/mantis
Additionally, it describes using Docker, gVisor runtimes, and relevant cloud SDKs, with sample daemon.json configuration for runsc runtime.
Recent releases
The Releases section states: "latest 0: - none", indicating no listed releases.
Traction
Stars: 480. Stars gained in last 24h: 9. Forks: 49. Open issues: 1.
Behind the repo
No linked startup or company information is provided beyond Google as the organization in the repo path.
Caveats
The README includes prominent warnings about running the suite in isolated environments due to potentially unstable or unsafe code execution. It emphasizes responsible use, human verification of findings, and limiting impact to non-production systems. It also states this is not officially supported by Google and is for demonstration purposes only. License is Apache-2.0. Created date is 2026-06-15; last push is 2026-07-21.