NEMESIS — Autonomous frontier red-team

Code

FAM-NEMESIS

Status

experimental — code maturity

Availability

lab-only — whether it is offerable

Ingest

none — how a target enters

Rules of Engagement

Required (acts on a live target)

Kind

Family (a selectable assessment)

Whether an autonomous agent can chain a multi-step compromise — judged step by step by a Tribunal against hidden ground truth, counted only if proven.

NEMESIS is RAVEN inverted — the same frontier-agent engine aimed at getting in rather than getting out — with a different oracle: a Tribunal that holds hidden ground truth the agent can never read. A crossing between stages is VERIFIED only when the agent presents the exact precondition the ground-truth edge requires, so it discovers the destination from the verdict rather than holding the answer key.

Its action space is synthetic: every action maps to a lab move, and even its “live” mode runs read-only commands against synthetic state in a fixture. Outcomes like “containment escape” are labelled simulated, and the Tribunal is configured so it mints no findings — a lab verdict is never a reported finding.

In the cyber-range, NEMESIS reasons a path “reach service A → obtain token → pivot to B”; each hop is adjudicated against ground truth and added to a capability graph only when proven. The same loop is never pointed at a real target.