PentAGI lab — self-hosted multi-agent security testing in an isolated lab
September–October 2026 (running)My own lab, as a playground for autonomous agents.
A self-hosted open-source environment (PentAGI) in which an AI system carries out security testing largely autonomously. A single orchestrator splits an assignment into subtasks and distributes them across specialised agents — planning, execution, refinement, enrichment, and a dedicated memory agent backed by vector storage. The agents work in an isolated lab: deliberately vulnerable images that I own (including Metasploitable), separated from my real network by an explicit firewall rule. The findings are not the story; the setup is the story.
- Docker (self-hosted)
- PentAGI (open source)
- Specialised agents with shared vector memory
- Kali Linux container as the executor
- Local models via Ollama, cloud models as fallback
Technically interesting
- Multiple specialised agents behind one orchestrator, with a separate memory agent so knowledge is shared between runs and between agents.
- Local-first. The default engine runs on Ollama; the token statistics show most usage ran locally or via free endpoints, with paid cloud models only as fallback.
- Real execution, not simulation: several hundred to over a thousand tool calls per run, the overwhelming majority terminal commands inside the Kali container; some runs went on for 6+ hours.
- Honest reporting: a partially successful exploitation was recorded by the system itself as “failed” because the requested proof of access was missing — it reports partial failure rather than claiming success.
- Operator-in-the-loop between runs: where a run stalled on network access, a manual firewall change and a restart of the target resolved it; the logs show exactly where the human was needed.
- Scope discipline as an observation: asked for a “quick scan”, the agent delivered considerably more than requested. Great for the client — and it underlines why a lab of deliberately vulnerable images is the right playground.
StatusRunning; six completed runs against three lab images in two weeks.