· 6 min read
PentestGPT is one of the projects that made LLM-assisted security testing credible. It is an open-source, interactive assistant that keeps track of a test's state and tells a human operator what to try next, which command to run, how to interpret the output, where to pivot. If you want a knowledgeable second brain while you drive the engagement yourself, it is genuinely good at that, and this article is not a knock on it. It is for people who have hit the one edge that defines it: you are still the one running every command.
The real difference: advisor versus operator
PentestGPT sits next to you. It reasons about the test and produces guidance; you execute. Darkmoon is built to run the engagement itself. The model plans and decides, but it does not get a shell. Instead it selects tools through a controlled Model Context Protocol layer, and a toolbox actually invokes the binaries under a bounded executor. That is the difference between a tool that tells you what to type and one that carries the chain end to end and hands you a report. We describe that architecture in multi-agent AI pentesting without giving the model a shell.
Where PentestGPT stops and Darkmoon continues
| Dimension | PentestGPT | Darkmoon |
|---|---|---|
| Role | Advises a human, who executes | Runs the engagement through a controlled tool layer |
| Model location | Built around cloud LLM providers | Local model by default (Ollama or llama.cpp), cloud optional |
| Sensitive data | Prompts and output go to the model as-is | Privacy gateway sends deterministic placeholders, not real IPs, hosts or credentials |
| Scope | Guides classic web and host testing | Adds first-class Active Directory and Kubernetes attack paths |
| Output | Interactive guidance for the operator | Findings graded by demonstrated impact, with the payload and raw response |
What that buys you
- Coverage and cadence. An autonomous operator re-runs the same disciplined chains across every host and identity without tiring, which a human-in-the-loop advisor cannot do at scale.
- Data that stays yours. Running on a local, self-hosted model plus the privacy gateway means the model works from markers, not your production values.
- Proof, not a to-do list. Each finding is marked EXPLOITED, CONFIRMED or capped low when only a lead exists, so you get evidence rather than suggestions to chase.
When PentestGPT is still the right pick
If you want a hands-on assistant that teaches as it goes, keeps you fully in control of every action, and you are comfortable executing each step yourself against a cloud model, PentestGPT does that well and is a fine choice. Darkmoon is for the case where you want the model to actually do the work, locally, with the data controls and the AD and Kubernetes reach that a guided assistant does not aim to provide. Both are open source; they are simply built for different jobs.
Full disclosure: Darkmoon is our project (GPL-3.0), self hosted and auditable. Read every line: github.com/ASCIT31/Dark-Moon, or start with the field guide to open source AI pentest tools.