SENTRY is an autonomous AI penetration tester that proves vulnerabilities with working exploits, not scanner noise. It runs entirely in your environment, and it physically cannot test a target outside a scope you have declared, attested and had approved. UK/EU-hostable, no US CLOUD Act exposure.
An autonomous offensive tool is only responsible if it cannot be pointed at a target you have not proven you may test. SENTRY enforces that in code, not in a prompt.
Every network-touching probe passes an in-code allowlist first. A target is permitted only if it belongs to an engagement you have declared, attested authorization for, and a human has approved. No attestation, no scan.
Every finding is reproduced with a second confirming request before it is filed, and re-checked adversarially — cutting the 40–70% false positives raw scanners produce.
Runs in your environment — on-prem, air-gapped, UK/EU. Your targets, findings and traffic never leave the box. Swap the reasoning brain for local inference and nothing phones home.
SENTRY is not anonymous self-serve. You cannot point it at an arbitrary URL. Access is granted per-engagement, the way every responsible offensive-security platform works.
On an authorized run against a staging secure-messaging relay, SENTRY confirmed these with working PoCs — 0 false-positive criticals, every auth gate and the crypto core independently verified sound. Pointed at a separate deliberately-vulnerable API, the same engine reproduced a real password-leaking endpoint and missing-control findings — each confirmed by an independent second request before filing.
Unauthenticated read of a victim's inbox metadata — sender graph, timing, volume. Content stayed encrypted; the metadata didn't.
An endpoint returned populated password fields in its response body. Reproduced, then reported with the exact fix.
Offline P-256 ECIES / AES-GCM / ECDSA harness — AEAD tamper, invalid-curve, cross-recipient and forward-secrecy properties, the QA a scanner cannot do.
The engine costs £0 to run, so continuous per-commit testing doesn't burn a per-test budget. Access is sold per authorized engagement — purchase does not grant instant, unattested scanning.
No — by design. You declare a scope and attest authorization, a human approves it, and only then can the engine touch those exact hosts. This is the same model responsible platforms (Cobalt, HackerOne, Intruder) use, and it is why SENTRY is safe to run autonomously.
No. The engine, the target and the findings run locally. The only optional outbound call is to the reasoning brain — point that at a local model for a fully offline, air-gapped deployment with zero external calls.
Scanners flag potential issues and produce 40–70% false positives. SENTRY reproduces each finding with a working exploit and has an independent model try to refute it before reporting — so what you get is validated and actionable.
An in-code allowlist (not a prompt instruction), fail-closed by default, plus a kill-switch. A target is permitted only if it is an approved, attested engagement scope. Every attempt, permitted or refused, is logged.