Skip to the content
AI and Cybersecurity Learning and Practice Center Test automation Open the Learning Center

NewsResearch

Booz Allen: frontier AI agents with no OT knowledge met all eight attack objectives in its lab, and moved a robotic arm within minutes

In Booz Allen's isolated OT lab, a mock industrial network with PLCs, HMIs, SCADA and a robotic arm, frontier models given no source code, documents or OT guidance ran eight attack scenarios and met every objective, The Register reported on October 11 and SecurityWeek on October 14 after the findings were presented at SecurityWeek's ICS Cybersecurity Conference in Nashville. One run went from a perimeter compromise to actions inside the control network in just over 16 minutes; another found the robotic arm's interface, mapped its protection zones and motion limits and moved it. The models chained asset discovery, device research and attack planning that normally takes several people, adapted when blocked, and no single safeguard consistently stopped them. Booz Allen's advice is to stop trying to predict every form of attack and instead make environments harder to traverse, remove implicit trust, detect consequential industrial changes fast and limit what an intruder can affect. Dark Reading's coverage on October 16 adds Claroty's account of a coding agent chaining two known refrigeration-controller flaws to root in 28 minutes for about six dollars of tokens.

Why it matters

Security through obscurity was the last line for a lot of industrial networks; an agent that reads the device manuals on the way in does not need to know the plant. Segmentation and change detection are what held, and only partly.

Sources

Read it with the rest of the news in the Learning Center