mantis
An AI agent toolkit to autonomously find, reproduce, and patch vulnerabilities in codebases.
Modular, stack-agnostic toolkit for AI coding agents to find, reproduce, and patch code vulnerabilities.
Mantis is a portable toolkit providing skills and an ADK reference harness for AI coding agents to autonomously handle security reviews, threat modeling, and patching. It is designed for security professionals and developers working in isolated environments who want to orchestrate AI models for vulnerability research.
What it is
The toolkit loads skills on demand—from threat modeling and research to code review, patching, and reporting—using an ADK reference harness to run pipelines. It builds semantic code indices, creates research hypotheses, statically verifies production viability, attempts exploit reproduction, and uses an adversarial loop to verify patches.
Who it's for
It suits security engineers and developers comfortable running autonomous AI agents and sandboxed pipelines for deep code audits. It does not suit anyone looking for a turnkey, production-ready CI/CD blocking gate out of the box, as the pull request scan features remain untested against live projects.
Key features
Autonomous Threat Modeling
Automatically generates threat models and builds semantic indexes for efficient code navigation.
Exploit Reproduction
Reproduces vulnerabilities through static guesses, unit tests, or spinning up mock servers.
Adversarial Patching Loop
Patches discovered vulnerabilities and runs an adversarial verification loop to ensure fixes hold.
Research Graph Synthesis
Designs custom agent-graph topologies on the fly based on specific plain-language objectives.
Up and running in 3 steps
Our honest verdict
What's great
- Includes modular skills for the entire lifecycle from threat modeling to patching
- Supports SQLite-backed knowledge persistence across multiple parallel runs
- Allows dynamic research graph synthesis for custom security review objectives
- Features risk calibration rubrics to combat LLM severity inflation
Where it falls short
- Requires strict isolation in a sandbox to prevent unexpected destructive actions by autonomous agents
- Pull request scans and associated GitHub actions are completely untested end-to-end
- Demands careful manual verification of all AI-generated findings by a human security expert
- Token spend and API costs can accumulate rapidly during unattended parallel scanning campaigns
Pricing & license
Mantis is free under its open-source license, but running it incurs API costs for frontier language models like Gemini depending on your usage volume and token consumption.
Open remote AI-training roles
880 open remote AI-training roles, updated November 15, 2026; most pay $20–$175 an hour. Zekai lists them free; the application goes to the employer’s platform.
Alternatives
These reviewed AI tools do the same job instantly — from our scored directory.
Use it, or have us set it up
FAQ
Can I run Mantis on production systems?
No. Mantis executes autonomously generated code that may be unstable or perform unexpected actions, and must only be run in isolated, restricted environments.
Are the AI findings guaranteed to be correct?
No. AI models are non-deterministic and can hallucinate findings or generate incorrect patches, meaning all results must be manually verified by a security expert.
How does Mantis handle budget limits during long scans?
You can set a hard token spend ceiling using flags like --max-tokens or run campaigns without a budget ceiling using --no-budget.
