farhanashrafdev/mantis — reverse-engineered prompt

Build me a command line tool for red teaming LLM apps so I can point it at a chat API and quickly see if it has obvious AI security issues. I want it to run a bunch of automated attack prompts against a target endpoint and check for things like prompt injection, hidden prompt or secret leakage, hallucinated links or facts, and basic tool or agent abuse. It should be easy to run with one command against a URL, but also support a simple config file for things like headers, auth token, request and response fields, retries, timeouts, and which test modules to include or skip.

Please make the results readable in the terminal, and also export JSON and SARIF so it can be used in CI and uploaded to security tooling. Add severity scoring for findings, map checks to the OWASP LLM Top 10 where it makes sense, and use exit codes so a pipeline can fail on serious issues. If helpful, include a small sample app and Docker support. Look up current docs online if you need to.