An operator
with a leash
"AI security testing" usually means a scanner with a chatbot bolted on and nobody watching. Ours is a bounded tool-use loop supervised by one of our security testers: the model reasons about your surface and chooses what to look at next, but every tool call passes a scope guard, every intrusive action waits for a human, and the whole run is written to the audit trail.
Seven rules, none of them optional
These aren't settings. They're enforced server-side on every tool call, and the shipped default refuses more than it allows.
Authorisation first
No run exists without a scope you have signed. An empty scope is refused outright — there is no "just have a quick look" path.
roe requiredScope allowlist
Domains, IPs and CIDR ranges are matched server-side on every single call. Anything outside the list is blocked and the block is logged.
enforced in codePassive by default
Active tooling only exists if a scanner allowlist is configured. Leave it empty and the engine is observation-only.
opt inApproval gate
Every active action calls a human-in-the-loop gate before it runs. The bundled gate denies by default, so the system is safe until you deliberately wire approval.
deny by defaultTime-boxed
Runs outside the agreed window are blocked, even mid-test. The window is part of the scope, not a calendar reminder.
windowAllowlisted binaries only
Configured scanner binaries, arguments passed as a list with no shell interpretation. It is not an exploit runner and cannot be turned into one.
no shellBounded and logged
Turn and token budgets cap every run. Start, every tool call, every block, every finding and the finish are written to the audit trail — and API keys never enter a prompt or a log.
full transcriptSkills, not a fixed checklist
The engine loads offensive-security skills — reconnaissance, web, identity — and composes them into a system prompt with a safety preamble. The model then works your objective inside the scope, following what it finds rather than marching through a list in the same order on every target.
- Adaptive probing that follows the evidence
- Each issue confirmed before it reaches the report
- Findings raised through a single structured tool, never free text
- A clean pass is recorded as evidence toward the control
- Human sign-off before any report reaches a client
Someone is always watching. The loop is bounded in code, but it is also supervised in practice — a tester follows the run as it happens, answers the approval gate, and pulls it up if anything looks off. The guardrails above are the floor, not the whole safety story.
Where the findings go. Each validated issue becomes a finding tagged AI Validation, linked to its Annex A control — so a discovered vulnerability lands as a nonconformity in the same register your auditor already reviews, rather than in a PDF nobody opens again.
A faster, tireless first pass
It covers ground quickly, never gets bored on the eightieth host, and writes up what it finds in a consistent format that maps straight into your control register.
- Broad coverage inside the authorised scope
- Consistent, structured write-ups
- Evidence captured as it goes
- Runs again cheaply for a re-scan
Not a replacement for judgement
We're direct about this: a report going out unreviewed is a liability. A qualified tester scopes every run, monitors it while it works, triages what comes back and signs the report. The engine is a force multiplier for them, not a substitute.
- Scoped and supervised by a security tester, start to finish
- Runs are monitored live and can be stopped at any point
- Active actions need explicit human approval
- Business-logic flaws still want a human looking at them
- Reports are triaged, reviewed and signed by a person