RA-01
Agentic attack planning
How autonomous operators decompose an objective into staged attacks, chain findings as they develop, and decide when a route is exhausted. The planning formalism underneath SIEGE.
SIEGE
Working lab — pre-launch
Methodology first, engines second.
The engines are in development; the methodology underneath them is published as it settles. Research is the part of the lab that is open now — tracks, writing, and a disclosure policy you can hold us to.
//01
Research tracks
Five standing tracks. Each one feeds an engine or the evaluation of the engines — nothing here is ornamental.
RA-02
Scope enforcement and control
Rules of engagement as machine-checked contracts — boundary verification on every step, operator-held checkpoints, kill-switch semantics, and a complete audit trail of agent decisions.
ALL ENGINES
RA-03
Human-layer simulation
Pretext design, consent models and measurement for simulation aimed at people and process — without harvesting anything live. The methodology underneath PRETEXT.
PRETEXT
RA-04
Evidence engineering
Attack graphs and append-only action logs: what it takes for a machine-executed attack to count as reproducible evidence rather than anecdote.
SIEGE · PROBE
RA-05
Evaluation of autonomous offense
Benchmarks and scoring for systems that attack — honest baselines, adversarial robustness of the evaluation itself, and reporting that refuses theater.
LAB
An engine that cannot explain its attack did not attack. Explanation is not a feature of the product; it is the product.
//02
Writing
Working notes from the tracks above. Full texts publish here as they clear internal review — newest first.
| INDEX | TITLE | DATE | TAGS |
|---|---|---|---|
| W-06 | Benchmarking Vision-Language Models on Low-Resource Bengali Handwriting Recognition | 2026-10 | VLM EVAL BENCHMARK |
| W-05 | Attack graphs as evidence: reproducibility requirements for autonomous red teams | 2026-06 | EVIDENCE |
| W-04 | Reachability over severity: ranking external exposure by path to impact | 2026-04 | SURFACE RECON |
| W-03 | Consent boundaries in human-layer simulation | 2026-03 | HUMAN LAYER ETHICS |
| W-02 | Evaluating autonomous offense: benchmarks without theater | 2026-01 | EVALUATION |
| W-01 | Coordinated disclosure for findings an agent found | 2025-11 | DISCLOSURE |
//DRAFTS AVAILABLE ON REQUEST — [email protected]
//03
Coordinated disclosure
We build offensive systems; we are obliged to receive reports well. The policy is short, written down, and applies to our own infrastructure and engines.
//HOW TO REPORT
Report a vulnerability
Send reports to [email protected], encrypted on request. Good-faith research that respects the policy below — and stops at demonstration — receives safe harbor. We will not support legal action against a reporter who follows it.
Agent-generated findings follow the same discipline: every report we issue is reproducible, scoped, and coordinated under a stated window.
- CHANNEL[email protected]
- ACKNOWLEDGMENTwithin 48 hours — target
- COORDINATION WINDOW90 days from report
- PUBLICATIONcoordinated after fix, or at window close
- SAFE HARBORgood-faith research, within policy
- CREDIToptional, always offered
//04
From the lab
Short notes on how the work is done — published between papers, dated and unpolished on purpose.
2026-09 // LAB NOTE
Why we publish the methodology
An offensive capability that cannot be inspected should not be trusted to run. Publishing the methodology is the offer: read how the engines are designed to think — how scope is enforced, how evidence is captured, where the operator sits — and then judge the outputs by that standard. It is slower than selling. It is the only version of this that we would believe ourselves.
2026-08 // LAB NOTE
On honest numbers
Until pilots report, every number on this site is a design specification or a labeled target. When field results exist, they will be labeled as field results — with the sample size attached. Nothing else counts.
//ENGAGEMENTS
Request an engagementResearch collaboration is open now, independent of engine maturity — MODEL-03 on the products page.
//RESEARCH CONTACT
[email protected]Vulnerability reports and research correspondence — policy above, 90-day coordination window.