Break in, by the book.

Every engagement runs under written rules of engagement with a live kill switch. Autonomy is powerful — so we built the leash first.

01 — Authorization

Nothing happens without paperwork.

Before a single agent touches a target, we agree and sign a scope: which domains, apps, and accounts are fair game; what's explicitly out of bounds; testing windows; and who gets paged if something looks wrong. You'll prove control of in-scope assets — we don't take anyone's word, including yours.

The ROE is a living document. Expand it, shrink it, or suspend it mid-engagement — changes take effect within minutes, swarm-wide.

02 — Guardrails

What the swarm will and won't do.

Always on:

  • Rate limits honored per asset — probes pace themselves like a cautious human, not a scanner
  • Out-of-scope assets hard-blocked at the agent layer, not just filtered in reporting
  • Destructive actions disabled class-wide: no writes to production data, no persistence, no ransomware simulation without explicit opt-in
  • Every agent action logged and attributable

Requires explicit opt-in:

  • Social engineering and phishing simulations
  • Physical-adjacent testing (badge systems, wireless)
  • Data exfiltration demonstrations — we prove reach with canary records, never bulk real data, unless you ask
  • Denial-of-service class testing (rarely; usually a loud "yes, reachable" is proof enough)
03 — The kill switch

Fifteen minutes, any hour.

Every engagement includes a 24/7 operations channel. One message — "stop" — and all agent activity against your estate halts within fifteen minutes. A human operator confirms, and the pause holds until you say otherwise. Test it on day one; clients do.

# engagement control channel
ops#1847> stop
[00:01:12] swarm halt acknowledged — 14 agents parked
[00:04:30] operator elias confirms: all quiet.
04 — Your data

Evidence, minimized.

Findings carry the minimum evidence needed to prove the path — request/response captures, canary records, access demonstrations — encrypted in transit and at rest, access-logged, and purged on a schedule you choose at engagement end. We don't train shared models on your estate's data. Your telemetry improves your swarm profile only.

Our own house: see responsible disclosure and privacy. Found something? security.txt.