AI code scanners cry wolf too often, and Google built a new open-source framework to make them prove their claims. The company estimates that conventional model-powered scanning can return true-positive rates below 7 percent, with the rest of the alarms pointing at bugs that do not exist.
Mantis, released under the google/mantis repository on GitHub, hands the work to a team of agents with separate jobs. A strategist agent studies how the repository is built, which parts were patched before, and what threat model applies. Research agents then follow data flows, control flows, and sanitization logic through the source. Two more agents act as critic and reviewer, filtering the candidates until only plausible findings remain.
Anything that survives those checks has to be reproduced. Mantis spins up a sandboxed environment and attempts to trigger the flaw, turning a model’s hunch into demonstrated evidence. Google says the approach answers the field’s grounding problem, where scanners rely on an LLM’s judgment instead of proof. Summarizing analyzed files into a hierarchical tree also trims token use by about 85 percent while keeping the structural detail reviewers need.
The framework ships as a modular skill suite with more than 15 tools that run in sequence or in parallel. A getting-started guide on the Google Cloud blog this week walks outside teams through setup, and Google describes Mantis as part of its internal drive to find and fix vulnerabilities at machine speed. For security shops drowning in alert noise, the pitch is simple: agents that only ring the bell when they can show you the damage.
