T3MP3ST
A multi-agent offensive-security framework that turns an already-running AI coding agent into an automated tool for authorized penetration testing.
T3MP3ST runs a recon, exploit, and report chain against an authorized target from a browser interface or the CLI, driven by whichever AI coding agent the user is already signed into, including Claude Code, Codex, or a fully offline local model through Ollama, LM Studio, or vLLM, so it needs no new API keys or cloud account. The project reports a 90.1% pass rate on XBOW's 104-challenge benchmark suite, above XBOW's own reported figure, plus solves against CTF challenges without hints and a cold hunt against real post-cutoff CVEs the underlying model had not seen during training.
Every benchmark number in its documentation is designed to be recomputed from committed data with one verification command, and a status table states plainly which parts of the project are stable, experimental, or still on the roadmap.
It is written in TypeScript, distributed under the AGPL-3.0 license, and is intended only for systems the user owns or has explicit written authorization to test.