Google's Mantis open-sources AI-driven vulnerability discovery and remediation, cutting token overhead by more than 85%.
The AMW Read
Google extends its known agentic-security push (alongside CodeMender) with a concrete multi-agent verification and hierarchical-summarization technique, a segment-level devtool/agent update rather than a new player or debate-resolving event.
Google's Mantis open-sources AI-driven vulnerability discovery and remediation, cutting token overhead by more than 85%.
On September 2, 2026, Google published technical details on Mantis, an open-source framework that automates vulnerability discovery, reproduction, and remediation in software repositories. Google states that raw AI code scans can produce a true-positive rate below 7%, so Mantis chains discovery agents with separate review and critique agents that confirm exploit conditions, reproduce the issue in an isolated sandbox, and only then generate a fix. Before analysis, Mantis mines a repository's change history and code structure to auto-build a threat model, working even when developers have no existing security documentation. To manage cost on large codebases, it builds hierarchical security summaries at the file and directory level before assembling a repository-wide view, which Google says cuts token overhead by more than 85% while preserving structural detail. Mantis integrates with Gemini CLI and Antigravity CLI, and ships a companion mode, mantis-advise, that surfaces prior fixes during development. Google still recommends human verification of AI-generated reports and patches, and isolation from production systems.
The release formalizes a defensive counterpart to AI's now-demonstrated offensive capability in vulnerability discovery, and pairs it with a concrete engineering answer to the context-cost problem that has limited AI security review on large repositories. Multi-agent verification chains aimed at suppressing false positives, plus automatic threat-model generation, position code-security review as a native layer of the coding-agent stack rather than a separate scanning product bolted on afterward.
For builders, the layered discovery-critique-sandbox pattern is a usable template for cutting false-positive burden in any AI code-review pipeline. For investors, standalone AI vulnerability-scanning startups now compete against a free, Google-distributed baseline wired directly into the CLI tools developers already run, which raises the bar for differentiation on accuracy, remediation quality, or workflow depth.