THIRD LEVEL LABDETROIT · RESEARCHSign in
REAL HARDWARE. RECORDED ACTIONS.

When AI holds
the keys,
keep the evidence.

We run a cyber range on real enterprise hardware. AI red agents attack. AI blue agents defend. We preserve what they did, what the systems saw, and what a human concluded.

We are building toward a sealed incident bundle at the end of every round—one a third party can verify for themselves.

—
complete sealed bundlesChecking the publish directory…
01 / THE RECORD

A round ends with a record
we can examine.

A complete bundle brings together the objective, both agents’ transcripts, the signed chain of actions, and the human verdict. It carries the telemetry: Wazuh alerts, packet capture, firewall and switch logs, and configuration diffs, alongside configs from before and after the round.

We include a verifier anyone can run. Missing evidence stays missing; a signature does not make an incomplete record complete.

02 / HOW IT WORKS

Two sides.
One record.

We give each side a bounded objective in an isolated range. We record their actions through VIRP and retain what the collectors observe. We compare the agents’ accounts with those records and add a human verdict. We seal the evidence so another person can check it with the public verifier.

The evidence mapStatic illustration of red and blue agents, an audit-only broker, three range guests, and evidence collectors. This is not a live round.RED AGENTBLUE AGENTAUDIT-ONLYBROKERGUEST 01GUEST 02GUEST 03COLLECTORS → SEALED EVIDENCE
Static illustration · no live activity or sample results
03 / WHAT IT IS FOR

Incident corpora

We build records for AI SOC vendors and labs to study how agents investigate, respond and fail.

Evidence for insurers

We collect data on what happens when agents hold admin rights, including the limits of the controls around them.

Operating doctrine

We use the record to improve how humans and agents share infrastructure: who acts, what they can change, and what must be checked.

04 / WHAT WE CAN CLAIM

Research, not a demo.

We are building an incident corpus, not a staged success story. Our completed capability demonstrations are not complete incident bundles, and we do not count them as such.

Same-operator witnessed. We operate the range and the witness. That gives us durability, not independent observation.

Blue actions were recorded, not gated. The audit record does not claim that a policy engine prevented unsafe actions.