Google Open-Sources Mantis Security Toolkit for AI Coding Agents

Google has open-sourced Mantis, a stack-agnostic toolkit designed to equip AI coding agents with a comprehensive set of security review skills, enabling them to manage the complete vulnerability lifecycle. This toolkit allows agents to identify suspected flaws, eliminate false positives, reproduce bugs within a secure sandbox environment, develop minimal code patches, re-test those patches against potential exploits, and finally, assign a risk score to any residual vulnerabilities. Mantis is not intended as a standalone scanner but rather as a collection of slash commands that can be integrated into existing AI coding agent frameworks. It operates under a strict set of rules governing code execution. While currently suitable for local and internal evaluations, Mantis is not yet ready for production deployment. Users can clone the toolkit and integrate it with agent frameworks such as Gemini CLI, Antigravity CLI, or the Google ADK. The Mantis pipeline structures its operations into distinct skill directories, each invoked as a slash command and executed sequentially. A central skill, /mantis-meta-agent, can orchestrate the entire vulnerability assessment and remediation loop within a continuous session. The initial stages focus on understanding the target system. The /mantis-history skill analyzes version control history for past security fixes, while /mantis-summarize generates directory maps. The /mantis-architecture skill constructs a Markdown knowledge base, and /mantis-threat-model derives trust boundaries. Subsequently, /mantis-plan creates a targeted roadmap for the security review. The middle stages are dedicated to identifying and refining potential vulnerabilities. The /mantis-researcher skill scans files based on the generated plan, followed by /mantis-dedupe, /mantis-review, and /mantis-critic, which collapse duplicate findings, enforce negative rules, and discard issues deemed impossible in a release build. The final stages involve proving and patching vulnerabilities. The /mantis-reproduce skill executes payloads in a gVisor environment or a virtual machine with disabled networking. The /mantis-chain skill constructs multi-step exploit chains from confirmed individual findings. The /mantis-patch skill applies and verifies code fixes. The /mantis-calibrate skill assigns a risk score ranging from 1 to 10. The /mantis-reflect skill records learnings to inform subsequent assessment passes, and /mantis-report generates a human-readable review document. A recently added skill, /mantis-advise, inverts this process by querying the accumulated threat model and past bug data to provide proactive security recommendations. This modular approach allows for flexibility and extensibility, enabling developers to customize the vulnerability assessment pipeline according to their specific needs and the characteristics of the target software.