In development — pilot stage
Three engines under one doctrine: autonomous agents that execute authorized attacks and document them like engineers. Each section below states the design as a specification — and its true status.
//01
ENG-01 // RED TEAM
SIEGE executes multi-stage attacks inside a written scope: initial access, credential chaining, lateral movement, escalation. One supervising human operator holds the rules of engagement and the kill-switch; three autonomous operators do the work, handing each other objectives as chains develop.
Every action is chosen because the previous one succeeded, and every action lands in an append-only log with its evidence attached. An operation reads back as an engineering postmortem — inputs, decisions, chains, exits — not a PDF dump. What is described here is the design; pilot engagements will report against it.
ATTACK PATH — TRAINING RUN
FINDINGS — TRAINING RUN
Build server holds a long-lived signing token the objective host accepts without replay checks.
Credential broker exposes an unthrottled verification endpoint — guessing goes unchallenged.
Staging panel discloses internal host names and component versions.
| PARAMETER | VALUE | STATUS |
|---|---|---|
| CLASS | Autonomous red team | Design |
| OPERATORS / RUN | 3 autonomous operator agents, 1 supervising operator | Design |
| SCOPE MODEL | Written rules of engagement, machine-enforced | Design |
| AUTONOMY LEVEL | Supervised; operator-held checkpoints at escalation | Design target |
| EVIDENCE | Append-only action log, evidence captured per step | Design |
| OUTPUT | Evidence chain, attack graph, remediation advisory | Design |
| AVAILABILITY | Pilot engagements from Q1 2027 | Planned |
siege spec --engine siege --dump
scope written roe · machine-enforced · kill-switch armed
operators 3 autonomous · 1 supervising
output evidence chain · attack graph · advisory
status in development — pilot Q1 2027
spec valid · envelope signed · illustrative
//02
ENG-02 // SURFACE RECON
PROBE indexes the external estate the way a recon pipeline would — domains, ranges, certificates, services, published code — then keeps watching. Exposure is ranked by reachability and path to impact, not by severity theater: what counts is what opens a route to something that matters.
Collection is passive and external; no credentials are required. The watch cycle is a planned 24-hour diff — when the surface changes, the change arrives with its path attached. In development; first scoped evaluations run under the pilot program.
| PARAMETER | VALUE | STATUS |
|---|---|---|
| CLASS | Attack-surface reconnaissance | Design |
| COLLECTION | Passive, external, credential-free | Design |
| SOURCES | DNS, certificate transparency, service mapping, published code | Design |
| WATCH CYCLE | 24-hour continuous diff | Planned |
| RANKING | Reachability-weighted, path-to-impact | Design |
| OUTPUT | Surface diff, exposure register | Design |
| AVAILABILITY | Scoped evaluations from Q1 2027 | Planned |
probe watch --scope ol-corp --diff
sources passive · dns · ct logs · published code
cycle 24h continuous diff — planned
ranking reachability-weighted · path-to-impact
status in development — evaluations Q1 2027
watch spec valid · nothing sent · illustrative
//03
ENG-03 // HUMAN LAYER
PRETEXT simulates the human layer: pretexts over email, voice and chat, run against named individuals inside a consented, written program. It tests process — how a helpdesk verifies identity, how an approval actually happens — rather than the click rate of a person.
Nothing live is harvested. Lures are drafted, checked against the rules of engagement, and delivered only through simulated infrastructure; findings describe gaps in process, not personal failures. In development; first exercises run as scoped evaluations.
Named contacts only. Simulated infrastructure. Nothing harvested.
Vendor renewal pretext, two variants, held for review.
“Your certificate expires today — approve the renewal here.”
Helpdesk asked for a callback number. None exists.
| PARAMETER | VALUE | STATUS |
|---|---|---|
| CLASS | Human-layer simulation | Design |
| CHANNELS | Email, voice, chat | Planned |
| SUBJECTS | Named individuals, consented programs only | Design |
| SAFEGUARDS | Simulated infrastructure; no credential or data harvest | Design |
| REVIEW | Every lure checked against rules of engagement before use | Design |
| OUTPUT | Process findings, human-layer advisory | Design |
| AVAILABILITY | Scoped evaluations from Q1 2027 | Planned |
pretext plan --program ol-human-layer
channels email · voice · chat — planned
subjects named · consented · written scope
safeguards simulated infra · nothing harvested
status in development — scoped evaluations
plan held for review · nothing sent · illustrative
//04
Three ways to put an engine to work, each stated at its true maturity. The engines are in development; these models describe how they leave the lab — and what is available today.
MODEL-01
PLANNED — Q1 2027First SIEGE runs inside a small number of authorized environments, engineered jointly with your team. Deep instrumentation, candid debriefs, terms set in writing. The cohort is limited by design — a pilot exists to answer questions, not to fill a calendar.
MODEL-02
IN DEVELOPMENTA fixed-scope PROBE or PRETEXT exercise against one defined surface, closing with a full evidence file: what was tested, what was found, what closes it. Narrower than a pilot and equally instrumented.
MODEL-03
OPENJoint work on agentic-offense methodology, evaluation and coordinated disclosure — the track the lab is built on. Available now, independent of engine maturity: papers, advisories and shared tooling under agreed terms.
//NEXT STEP
Every path ends in a written scope. Say which engine and which model interest you; we answer within one business day and say plainly what is ready and what is not.
The specification is the offer. Judge the engines by what they commit to — and hold the pilots to it.
