When your pentest agents stop taking orders and start taking initiative

What it does

Pentest Swarm AI runs autonomous penetration tests by coordinating specialist agents through a shared Postgres-backed blackboard. Recon agents feed findings with decaying “pheromone” weights; classification, exploitation, and reporting agents trigger themselves when the board state matches their predicates. It wraps eight ProjectDiscovery tools plus nmap, speaks Claude or any OpenAI-compatible model, and outputs markdown, HTML, JSON, or SARIF.

The interesting bit

The authors are almost more interested in taxonomy than exploits. Their core claim — backed by a detailed comparison table — is that every other “multi-agent” pentest tool is actually a pipeline with a central planner. They built stigmergy (indirect coordination via environment), emergence (unplanned attack chains), and decentralization (self-triggering agents) from scratch because, they argue, nobody else had. The pheromone decay is configurable per finding type: ports stay hot for hours, sessions for minutes.

Key highlights

• Real swarm architecture: no central orchestrator rewrites needed to add agents
• Scope enforcement at both tool and executor layers; cleanup registry runs on SIGINT, crash, or budget exhaustion
• Claude prompt caching enabled by default for recon and classifier to cut API costs
• --strict mode promotes LLM errors to fatal crashes rather than silent hallucinations
• AGPL 3.0; alpha status with a public implementation plan and honest feature-status labels

Caveats

• The stigmergic swarm scheduler itself is alpha; the stable default is still the sequential 5-phase runner
• Benchmark numbers are explicitly promised for a future release, not yet shipped
• Postgres blackboard backend is beta; the memory-backed version is what currently runs

Verdict

Security researchers and red-teamers who want to experiment with decentralized agent coordination should watch this closely. If you need a battle-tested autonomous pentest tool today, the sequential mode works but the swarm is still a research direction — albeit a well-documented one.