The SwarmHack story

SwarmHack™ takes its architecture from nature: many small specialists, sharing state, outperforming one large generalist. Instead of one model attempting an entire engagement, a swarm of focused agents each own a narrow job and hand their findings to the next.

Why a swarm beats a single model

Specialized agents keep context small and decisions verifiable, run in parallel across surfaces, and let a credential discovered on a web host become an Active Directory pivot seconds later through shared swarm state. A single large model asked to run a whole engagement loses the thread: context windows fill with noise, earlier discoveries fall out of memory, and the reasoning that connected step three to step eleven is gone by the time it matters.

From colony behaviour to a kill chain

In a colony, no individual holds the map. Coordination emerges from shared signals left in the environment. SwarmHack works the same way — agents write findings, credentials and host facts into shared engagement state, and a GOAP A* planner reads that state to select and order the next actions deterministically. When an action fails, the planner re-plans rather than aborting, which is exactly how a human tester behaves when an exploit does not land.

Proof as the organizing principle

The other inheritance from nature is ruthlessness about what counts. Every finding is graded Exploited, Observed, AttackPathIdentified or Simulated, and an "Exploited" claim without captured target output is downgraded automatically. The swarm is built to bring back evidence, not opinions.

See how the engine works, read swarm vs. a single LLM, or explore agentic pentesting.