SwarmHack: the attack execution engine behind Prancer

SwarmHack autonomously plans, executes, and chains authorized attacks across supported enterprise environments. Unlike a scanner that reports isolated vulnerabilities, SwarmHack uses the result of one successful attack step to determine and execute the next.

A scan finds a weakness. An attack changes state.

A scanner goes target → vulnerability found → report, and nothing about the environment has changed. SwarmHack goes target → exploit → new access → new information → new attack options → next action. When an exploit succeeds, a credential may now exist, a new host may become reachable, a new identity may become usable, or a security boundary may become attackable. SwarmHack carries that new state forward: the output of one attack becomes the input to the next.

Follow the attack, not just the vulnerability

The pivot graph shows why: initial internet access leads to a web application, an exploit yields a credential, the credential yields an identity, and the identity branches into cloud, VPN/ZTNA and internal application paths. One branch dead-ends and is rerouted; the others converge on an internal host, then lateral movement, then a segmentation boundary that is either blocked by the control or crossed to a critical asset. An attacker does not follow one predetermined line, and neither does SwarmHack. Every successful step changes the attack graph: a vulnerability becomes a credential, a credential becomes an identity, an identity creates new access, and new access creates another pivot.

Plan from the current attack state

After each attack step, SwarmHack updates its understanding of the environment and evaluates the authorized actions now available: current attack state → available actions → select next action → execute → observe result → update attack state, then repeat. It uses Goal-Oriented Action Planning (GOAP, implemented as GOAP A*) to reason across these attack states, so the plan changes as the attack succeeds or fails.

AI expands what SwarmHack can try

Known attack techniques are executed deterministically. Where extra reasoning helps, approved AI models explore possibilities, analyze unusual behavior, or propose new attack actions. The Prancer Model Router selects the appropriate intelligence source — deterministic logic, efficient model, frontier model, or private/local model — using capability, cost, latency, privacy and policy, and outputs an attack hypothesis for SwarmHack execution. Use AI where it adds value; don't require it where it doesn't. See Secure AI Factory.

SwarmHack does the attack

SwarmHack is not generating recommendations or attack descriptions — it interacts with supported systems to carry out authorized attack actions across application and API, identity, network and cloud, security controls, AI/agentic, and robotics/IoT domains. 85 registered plugins and 102 agent capabilities; detailed coverage lives on the Product and External Attack Exposure pages.

Execution becomes proof

SwarmHack executes → the target responds → evidence is captured → verdict: Proven (the attack worked and captured output is in the report), Control-blocked (the control stopped the path, recorded with evidence), or Unreachable (the path could not be completed from the tested starting point within authorized scope). A model prediction or attack hypothesis is never treated as successful exploitation — the target environment determines the result. If it isn't proven, it doesn't publish. See Evidence Architecture.

Evidence ships as OCSF 1.1.0 JSON, interactive HTML attack graphs, Markdown, DOT/Cytoscape/Neo4j, MITRE ATT&CK mapping and CVSS v3.1 scores, with rollups for PCI DSS 4.0, NIST CSF 2.0, OWASP Top 10 2021, SOC 2, HIPAA, ISO 27001:2022, DORA and NIS2.

Once you prove the path, you can test it again

Attack path proven → remediate → re-run same attack → blocked / unreachable. After a firewall rule, identity policy, code fix, segmentation change or other remediation, Prancer reruns the attack to determine whether the path has actually been closed, so a pentest finding can become a regression test. Structured execution makes that possible: byte-identical plans across runs, one static binary, ~8,800 tests across ~290 suites, and reports diffed against hand-authored ground-truth inventories.

Autonomous does not mean uncontrolled

Control: authorized scope with named targets and no wildcard drift, signed engagement authorization with ≤24-hour expiry, execution boundaries re-checked before every new target, destructive actions default-deny, a fail-closed process-global kill switch, and auditability of what was attempted. Deployment: SaaS, private cloud, on-premises, sovereign, air-gapped, or Secure AI Factory, from a single static binary for Linux x86_64/arm64, macOS x86_64/arm64 and Windows x64. SwarmHack is designed to execute inside explicitly authorized boundaries and can operate where security data cannot leave the customer infrastructure.

Proven in controlled live-fire testing

Controlled live-fire environment — not customer production. A fully instrumented AWS lab of 200 hosts across five isolated network tiers with seeded ground truth and cloud-native detections enabled produced a four-hop chain: unauthenticated web RCE with marker round-trip proof → service credentials captured in full → host root proved by a bounded read → cloud IAM credential holder via IMDSv2, Secrets Manager and sts:AssumeRole, reaching sensitive data. 18 application surfaces, 45 findings, 19 Exploited, 13 Critical, 66 crown jewels, zero cloud-native attack detections. Ruled-out branches were recorded with reasons: API blocked, legacy database not exploitable, Active Directory/Entra unreachable. Read the research.

One attack engine, many validation problems

Segmentation, Zero Trust, SASE/SSE, frontier attack validation, agentic identity & NHI, applications/APIs, identity, cloud/network and robotics/IoT are all validated by the same engine — SwarmHack — reporting from the same evidence architecture. Prancer is one technology platform, not a collection of unrelated products. Request a demo.