Google has open-sourced Mantis, an AI-agent framework built to automate the full software vulnerability lifecycle, from detection and validation through reproduction and patching. The tech giant created Mantis to tackle the persistent problem of false positives and fabricated vulnerabilities that plague traditional AI-powered code scanning tools. While careless AI code scanning commonly generates imaginary bugs and delivers true-positive rates below 7%, Mantis was engineered to perform better by merging standard agentic methods like critic and review agents with sandboxed vulnerability reproduction for verification.

Rather than scanning files through brute force, Mantis examines repository history, past security patches, architecture, and threat models. The framework condenses analyzed files into a hierarchical tree that captures directory and repository-level context, slashing token consumption by 85% while preserving crucial structural details. The system deploys industry-standard agentic techniques—including critic and reviewer agents—to weed out false positives and elevate significant discoveries. It can also replicate findings in a sandboxed setting, offering proof that a discovery represents a genuine vulnerability in a secure and managed manner, instead of depending exclusively on the language model's assessment.

According to Google, Mantis forms part of the company's internal strategy to locate and repair vulnerabilities at machine speed. Beyond critic and reviewer agents, the framework uses a strategist agent to assess high-level code structure, threat models, and dependency graphs, alongside research agents that leverage internal code searches to inspect raw source files thoroughly, tracking data flows, control flows, and sanitization logic. Mantis operates as a modular skill suite featuring more than 15 tools that can run sequentially or simultaneously, including mantis-summarize, mantis-review, mantis-critic, and additional capabilities. The stages exchange information by reading and writing to a shared state saved on disk, and the system accommodates multiple models as well as blending different models for distinct phases.

Google emphasizes that vulnerability scanning systems inevitably generate false positives, sometimes in a "frustrating number." Mantis tackles this challenge with the mantis-review stage, which uses a rule-based negative filter to remove probable false positives. However, the company stresses that the filter requires careful handling: low-risk discoveries shouldn't automatically be labeled false positives, since an excessively broad negative filter might weaken the system's capacity to spot authentic vulnerabilities. Google recommends deploying a "flash" or "lite" model variant for tasks lacking logic depth, such as quick classification tasks using mantis-researcher or grouping similar text patterns with mantis-dedupe, while more powerful models suit tasks demanding deep contextual comprehension and zero-shot problem solving, such as mantis-reproduce, which creates functional crash reproducers, and mantis-patch, which produces side-effect-free fixes for the codebase. The framework is now available on GitHub, with a detailed breakdown of all available stages, inter-stage contracts, and best practices provided in the agent reference guide. For organizations struggling with alert fatigue from conventional scanning tools, Mantis offers a path toward automation that relies on evidence rather than guesswork, potentially shifting security teams from triage mode to strategic remediation. The modular architecture also signals a broader trend in enterprise AI: specialized agents coordinating through shared state may prove more reliable than monolithic models attempting every task alone.