TLDRocket
Sign in

Google Open-Sources Mantis: A Modular Skills Toolkit That Lets Coding Agents Find, Reproduce and Patch Vulnerabilities

MarkTechPost Michal Sutter

Google open-sourced Mantis, a modular security toolkit for coding agents. It finds bugs, reproduces them, patches them, then re-checks the fix inside a sandbox.

Based on reporting by MarkTechPost, Michal Sutter — read the original for the full story.

Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error

Google has open-sourced Mantis, a modular security toolkit built for coding agents that already exist. It is not a point-and-shoot scanner. Instead, it drops in as a set of slash commands and rules that guide an agent through a full security review, from finding a suspected flaw to reproducing it, patching it, and testing the fix again.

That pipeline is split into separate skills. Mantis starts by learning the target with commands like /mantis-history, /mantis-summarize, /mantis-architecture, /mantis-threat-model and /mantis-plan. Then it narrows the field with /mantis-researcher, /mantis-dedupe, /mantis-review and /mantis-critic, which are meant to remove duplicates, apply negative rules and discard issues that could not happen in a release build.

The more interesting part comes later. /mantis-reproduce runs payloads in gVisor or a VM with networking turned off. /mantis-chain can combine confirmed findings into multi-step exploit chains. /mantis-patch applies and verifies the fix, /mantis-calibrate gives the risk a score from 1 to 10, and /mantis-report turns the whole thing into something a human can read. There is also /mantis-reflect, which feeds learnings back into the next pass.

Google says the model is aimed at local and internal evaluation, not production. That caution makes sense, because Mantis is built around a trust boundary that most agentic security tools skip: the reproducer and the re-attack. The company also says its hierarchical summary tree cuts token overhead by more than 85 percent, and it cites sub-7 percent true-positive rates for naive AI code scanning as the problem this setup is meant to address.

A newer command, /mantis-advise, flips the process around and checks the accumulated threat model, bug history and verified patch patterns before code gets written. That is the useful bit here. Fewer heroic prompts, more boring guardrails. Exactly the sort of unglamorous discipline security tooling usually needs.

My take — AI-written commentary, not fact-checked reporting

Mantis is the right kind of unsexy. It treats AI like a junior operator that needs rails, not like a wizard that can freestyle its way through security. That is a healthier bet than yet another glossy demo that confuses confidence with correctness.

Read more about this at: MarkTechPost

Related stories

The daily briefing

Every AI story that matters, in your inbox by 8am.

TLDRocket reads all relevant sources, removes duplicate coverage, and summarises the day in two minutes. Follow companies and topics for alerts, or get the briefing in Slack. Free, no spam, unsubscribe anytime.