For self-hosted & sovereign environments

Local-LLM pentesting that never leaves

Most autonomous AI pentesters send your code, IPs and traffic to a hosted model. Darkmoon can run entirely on a local LLM, with a Privacy Gateway that keeps sensitive values away from the model — so it is safe to point at production or a client environment.

Clone on GitHub How the Privacy Gateway works

Your wedge

The message no leader owns

Local model, tokenized data, readable source — the wedge the high-star cloud tools do not have.

Local LLM by default

Run against Ollama or llama.cpp on your own hardware. The 57-finding OWASP Juice Shop benchmark was produced black-box on a local model — proof the local path works.

The model never sees real values

The Privacy Gateway tokenizes IPs, hosts and credentials before the prompt reaches the model, then rehydrates the report server-side. On one measured run the model saw the target 0 times across 3.5 MB of traffic.

GPL-3.0 and auditable

Self hosted via docker-compose. Read and modify every line — no closed cloud in the loop, nothing to take on trust.

What you get

What sovereignty means here

Nothing leaves your perimeter
With a local model and the Privacy Gateway, no source, no traffic and no credentials are shipped to a third party. This is a design guarantee you can read in the code.
Breadth in one tool
Web, API, Active Directory & identity, Kubernetes, cloud, CI/CD, databases, IoT/firmware and LLM endpoints — 50+ specialist agents, all running on your infrastructure.
Reproducible, honest proof
The offensive benchmark and the 42/57 remediation dossier publish their limits and invite you to reproduce or contest them, black-box, on a public lab.

Start here