What SwarmHack is

SwarmHack is Prancer's AI-Native autonomous penetration-testing swarm: a Rust binary (swarmhack) that drives a coordinated multi-agent kill chain against an operator-supplied target, extracts crown jewels from each confirmed vulnerability, and emits an OCSF 1.1.0 Vulnerability Finding report (class_uid=2001, category_uid=2) plus customer-facing Markdown and HTML deliverables.

The three agent families

Agents are organised into three families orchestrated by a GOAP A\* planner over a kill-chain phase routing layer:

| Family | Count | Focus |

| --- | --- | --- |

| Web / API exploit | 23 | SQLi, XSS, CSRF, IDOR, SSRF, JWT, deserialization, etc. |

| AD / identity / network | 34 | LDAP, Kerberos, SMB, NTLM relay, DCSync, ADCS, Entra |

| Secure Service Edge (SSE) | 16 | ZTNA, DNS, SWG, CASB, DLP, RBI, GenAI, observability |

Total: 100+ customer-facing agents.

Optional model-assisted exploration

SwarmHack combines a deterministic execution core with an optional model-assisted exploration layer. The core uses versioned playbooks and GOAP A\* planning, requires no model, and supports fully air-gapped operation. When enabled, eligible frontier, private, open-weight, or sovereign models can broaden reconnaissance, generate attack-path hypotheses, and help prioritize what the swarm should try next.

The model never decides what counts as proven. Every proposed path returns to deterministic execution, and the same evidence gate applies in every deployment mode: captured target output is required before a finding is reported as Exploited or Critical.

What SwarmHack is not

SwarmHack is not a configuration auditor, not a Breach and Attack Simulation (BAS) platform, and not a vulnerability scanner.

The unit of output is an *exploit-validated* finding: every Critical / High finding in a report has a deep-exploitation evidence trail — extracted credentials, served files, dumped tables, replayed tickets — recorded in the same JSON. Scans that find nothing emit zero findings; SwarmHack never fabricates "informational" volume.

Reference metadata

| | |

| --- | --- |

| Binary version (Cargo.toml) | 2.6.0 |

| npm package (release/npm/package.json) | [email protected] |

| Operator audience | Prancer engineer, partner SE, customer security operator |

| Evaluator audience | CISO, security architect, partner technical lead |

| Companion docs | npm README, exploit-validated-sse-assessment whitepaper, SSE lab README |

[!IMPORTANT]

SwarmHack is not a publicly available tool. It is distributed to clients under signed contracts and asset-ownership documentation. Every engagement requires the operator to load a per-engagement authorization document via --authz-doc. See the Authorization Model for the safety contract.