Talos diverts hostile OT agent requests onto a plant twin
Original title: Stinging the Swarm: Extracting Attacker Intelligence from a Twin-Controller OT Honeypot
Across 2,400 autonomous attacker agents in four concealment tiers, no write reached the real plant.
Talos is an inline gate that judges language-model agent requests to industrial controllers in milliseconds and sends flagged agents to a device-faithful mirror of the plant. On 2,400 agents in four hiding tiers, the damage path was denied and a device-level envelope check backstopped it. The mirror also collects behavioral intelligence, including toolchain fingerprints and targeted registers, from agents ordered to stay silent. The gate, mirror, and harness are released under Apache 2.0.
Unconfirmed: Reported attacker identities are cover stories the agents authored, not independently confirmed.
- Tiers escalate from no concealment to an explicit non-disclosure rule and a trained cover story.
- Unmapped registers are denied without exception, while safe-direction stop commands remain allowed.
- Decisions were model-driven in every tier, with no inference timeouts.
- Only the paranoid, low-yield attacker region resists both capture channels.
- Identity fields are self-authored cover stories rather than verified identities.
From the original article
1 min read at arxiv.orgAutonomous language-model agents can drive industrial controllers over fieldbus protocols, probing a plant at machine speed and in coordinated swarms. We present Talos, an inline gate that judges every agent request to a controller in milliseconds and routes agents it flags as hostile to a device-faithful decoy -- a mirror twin of the real plant -- where their attack continues harmlessly and is read for intelligence.
By Canyon Grey Dell'Omo · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.