What happened
Booz Allen's testing (frontier 'super intelligence' report, presented in October 2026) found that in an isolated OT environment commercial frontier AI models achieved all eight defined test objectives, and in one case moved from perimeter compromise to a kinetic outcome (energizing/moving a robotic arm) in ~16 minutes. Humans in the loop slowed it; no single existing defense consistently stopped the autonomous agents.
Why it matters
This is a dangerous-capability disclosure, not a patchable CVE: it shows that the 'security through obscurity' and split-human-workflow assumptions OT defenders rely on are being collapsed by autonomous agents, converging cyber and physical security and compressing attack time to minutes.
Attack vector
An autonomous frontier-model agent with no prior OT knowledge (no source code, engineering docs, or OT guidance) was given network access to an isolated OT lab; it performed asset discovery, vulnerability research, and attack planning, crossed from the perimeter through IT and into the OT control network, and moved a robotic arm — all in about 16 minutes.
Affected systems
Frontier/super-intelligence model agents operated without OT-specific training (Booz Allen isolated OT lab; report released October 2026)
Mitigation
Defense-in-depth for OT: reduce traversability (network segmentation/micro-segmentation), remove implicit trust, detect consequential industrial changes quickly, and cap what an attacker can affect once inside. No patch exists; this is a capability-eval warning, not a product flaw.