AI models reportedly complete all eight OT attack scenarios in a modeled factory

Controllers and a robot operated in a modeled facility

According to The Register's account of Booz Allen testing, the models mapped industrial assets, identified vulnerabilities, developed access paths, moved through production systems, manipulated controllers, altered connected equipment, compromised SCADA and HMI systems, and located and moved a robotic arm.

The tests used a modeled general-manufacturing environment containing equipment from multiple vendors. Booz Allen did not identify the two frontier models it tested.

The report does not provide a directly comparable earlier test result for the same models, environment, and eight scenarios.

The reported finding is that the tested models achieved objectives in all eight OT attack-chain scenarios in a modeled multi-vendor industrial environment. The result was obtained under laboratory guardrails, including human approval before exploitation or actions with physical effects.

Execution required human authorization

The reported tests connect AI-assisted reconnaissance and attack planning with controlled physical equipment actions in an OT setting. They do not establish a real-world infrastructure compromise, and the agents operated with human approval gates for exploitation and potentially physical actions.