Almost every autonomous AI pentester ships your source code, IPs, hostnames or traffic to a hosted model. For a lot of teams, regulated, sovereign, or testing a client environment, that is a non-starter. Here is how to run autonomous pentests entirely locally.
What "local" has to mean
- Local inference. The model runs on your hardware (Ollama / llama.cpp), not a vendor API.
- No leakage even if you use cloud. Sensitive values are tokenized locally with deterministic placeholders before any prompt leaves, so even a cloud model never sees real IPs, hosts or credentials.
- Self-hosted, auditable. Runs on your infrastructure via docker-compose; source is readable.
Darkmoon is built exactly this way
GPL-3.0, local-LLM capable, with a Privacy Gateway for the tokenization above, and coverage across web, API, Active Directory and Kubernetes. Proof it works locally: 57 vulnerabilities on OWASP Juice Shop in 28 minutes on a local model(reproducible benchmark).
Full disclosure: Darkmoon is our project. github.com/ASCIT31/Dark-Moon.
See the proof, not just the write-up: the Pro remediation benchmark (fixes retested against the exploit) · how Darkmoon compares to other AI pentest tools.
← All articles