In September 2026, Google released Mantis, an open-source, stack-agnostic toolkit designed to equip AI coding agents with a comprehensive suite of security review skills. Unlike conventional scanners, Mantis enables agents to manage the entire vulnerability lifecycle: identifying potential flaws, filtering false positives, reproducing bugs in isolated sandboxes, generating minimal patches, re-testing those patches against attacks, and assessing residual risk. The toolkit is not a standalone tool but a collection of slash commands that integrate seamlessly into existing agent frameworks such as Gemini CLI, Antigravity CLI, the Google ADK, or other comparable platforms. It is currently suitable for local and internal evaluations but not for production deployment.
The Pipeline: A Structured Approach to Security
Mantis organizes its capabilities into a sequential pipeline, with each stage represented by a dedicated skill directory invoked as a slash command. A supervisor skill, /mantis-meta-agent, can orchestrate the entire workflow within a long-lived session, ensuring efficient coordination across stages.
Early Stages: Understanding the Target
Initial skills focus on building contextual awareness of the codebase. /mantis-history mines version control logs for past security fixes and patterns. /mantis-summarize generates directory maps for orientation, while /mantis-architecture constructs a detailed Markdown knowledge base of the systemβs structure. /mantis-threat-model derives trust boundaries and potential attack vectors, and /mantis-plan synthesizes these insights into a targeted, prioritized roadmap for vulnerability discovery and remediation.
Middle Stages: Discovery and Filtering
The core discovery phase begins with /mantis-researcher, which systematically scans files against the roadmap to identify candidate vulnerabilities. Subsequent skills, including /mantis-dedupe, /mantis-review, and others, refine these findings by removing duplicates, validating authenticity, and eliminating false positives. This multi-step filtering ensures that only genuine security issues proceed to the reproduction and patching stages, enhancing both accuracy and efficiency.
Later Stages: Reproduction, Patching, and Validation
Once potential vulnerabilities are confirmed, Mantis facilitates reproduction in a secure, sandboxed environment, allowing agents to observe the flaw in action without risking production systems. The agent then generates a minimal, targeted patch designed to address the root cause with minimal disruption. Finally, the toolkit re-executes attack scenarios against the patched code to verify its effectiveness, scoring the residual risk and providing a clear metric for security posture. This closed-loop approach not only accelerates vulnerability remediation but also helps prevent regression, making Mantis a valuable asset for DevSecOps teams in 2026 and beyond.
via MarkTechPost