Google Open-Sources Mantis Harness to Eliminate Hallucinated Vulnerabilities with AI Sandboxes

Google has made its internal vulnerability discovery and remediation framework, the Mantis harness, available as an open-source project on GitHub. The framework automates the end-to-end security analysis lifecycle: parsing repository history to extract past fixes, building dynamic architectural threat models, orchestrating multi-agent critics to evaluate potential vulnerabilities, reproducing candidate exploits inside isolated sandboxes, and generating verified patches. Mantis also introduces hierarchical security summary trees that reduce token consumption by over 85% across large codebases, alongside a developer-facing tool called mantis-advise to guide upstream code generation. Unconstrained LLM security scans are notorious for low true-positive rates—often falling below 7% due to hallucinated attack paths and irrelevant findings. When scanning tools output hundreds of theoretical bugs, security teams face alert fatigue and developers lose confidence in automated tooling. Mantis tackles this operational bottleneck by introducing strict grounding mechanisms. Rather than treating an LLM output as a verified bug, the system forces agentic workflows to validate hypotheses against an isolated execution sandbox before raising an alert. This changes the economics of automated application security, shifting the workload from manual human triage to automated proof-of-exploit verification. This launch aligns directly with the industry shift from passive Static Application Security Testing (SAST) tools toward active, agentic DevSecOps workflows. As software release velocity accelerates under AI-assisted development, vulnerability discovery must operate at machine speed without inflating false alarms. Traditional static scanners lack holistic architectural context, while generic frontier models often misinterpret codebase semantics. By structuring repository knowledge into hierarchical summaries and incorporating past commit history, Mantis models how human security researchers analyze systems: mapping the architecture, understanding domain context, and confirming exploitability before filing a report. For DevOps and platform engineers, operationalizing Mantis requires establishing robust, ephemeral sandboxing infrastructure within CI/CD pipelines. Security leaders should define explicit vulnerability acceptance criteria and seed Mantis with domain-specific threat context—such as non-critical crash boundaries—to prevent agents from chasing low-impact issues. Teams should also explore integrating mantis-advise directly into developer workflows to surface historical security lessons before code merges. The trade-off lies in compute overhead: running isolated reproduction environments and multi-agent loops increases compute consumption compared to basic AST scanners, making selective, risk-based scanning essential for high-throughput repositories.