An AI offensive security platform built on a deliberately starved model action surface.
Execution moves out of the model’s context, so every action can be gated, audited, or revoked. The reasoner is given structured facts, not shell access. An anchored, tamper-evident ledger keeps the model and the platform accountable.
Reasoning held apart from execution.
The model writes the request, and a malformed probe reads as a clean target. A tester who runs out of time writes a thin report; an agent writes a confident one.
The model picks the malformation; primitives do the counting. Judgment stays probabilistic, and correctness stays deterministic.
The target writes what the model reads and can try to steer it, so the head is bounded: nothing to exec, credentials out of view, undeclared effects refused.
Lengths, framing, HPACK and TLS state live in deterministic primitives, with controls inside every call. New primitives are model-written and operator-approved.
Intent goes down and facts come back, never raw responses. A dead probe is marked unusable before it’s compared, and silence is never scored as a negative.
Built for professional offensive security work, from scope to report.
Designed by a career operator around the ways engagements actually go wrong, with every part tied to the same anchored record.
capabilities referenceThe audit record protects the operator from the people they test for, and the people they test for from the operator.
Client, system owner or employer, whoever asks gets the same answer about what ran and when. Every ten minutes the chain heads anchor to Bitcoin, FreeTSA, DigiCert and Sigstore Rekor: four independent substrates, three separate failure modes.
how anchoring works · §09Anyone holding the record can verify it without us.
Our offline verifier walks the hash chain, replays per-event Merkle proofs and validates each anchor against pinned roots, with zero network calls to Disquiet.
how verification works · §09Bring your own reasoning client.
Disquiet provides the execution layer and the record. How much autonomy the reasoner gets is your workflow decision: one agent or several, ten minutes or ten hours, run in your own harness. We don’t take a side, and we won’t get in the way.
Provider keys stay yours and nothing pools. The reasoner reads structured facts, not raw responses, so the same work costs fewer tokens.
where the reasoner sits · §06