Google has open-sourced Mantis, a stack-agnostic toolkit of security review skills that enables AI coding agents to execute the entire vulnerability lifecycle. The system allows agents to identify suspected flaws, strip false positives, reproduce bugs in a sandbox, write minimal patches, re-attack those patches, and score residual risk.
Mantis operates as a set of slash commands for existing coding agents rather than a standalone scanner, chaining sequential skills through a supervisor agent. Key stages include mining version control history, building threat models, sweeping files against plans, and collapsing duplicates. The pipeline concludes with sandboxed reproduction using gVisor or VMs, assembling exploit chains, applying fixes, and generating human-readable review packets.
The toolkit addresses the problem of sub-7 percent true-positive rates in naive AI code scanning by grounding trust in sandboxed reproduction and patch re-attack rather than model confidence. Google reports that its hierarchical summary tree cuts token overhead by over 85 percent. While deployable locally under Apache 2.0 for evaluation, it is not yet recommended for production use.