Booz Allen Tests AI Agents on Unfamiliar OT Systems
Recent tests conducted in an isolated laboratory by Booz Allen demonstrate that frontier artificial intelligence models can successfully navigate unfamiliar industrial control systems and manipulate physical equipment in a matter of minutes.
Radwan AbdulhadiEditor and technology writer
Booz Allen Presents New OT Lab Findings
A new report examining the impact of artificial intelligence agents on operational technology was presented at SecurityWeek’s Industrial Control Systems (ICS) Cybersecurity Conference in Nashville. The findings demonstrate that frontier models can navigate unfamiliar industrial systems, adapt when blocked, and move physical equipment within minutes.
The study follows recent government terminology shifts, including a September 29, 2026 Executive Order addressing executive branch terminology. Observers note that the AI sector continues to grapple with a lack of consistent and formal definitions regarding terms like super intelligence versus artificial general intelligence.
Testing AI Agents in an Isolated OT Environment
To evaluate potential risks, Booz Allen examined the behavior of AI frontier models within its own isolated operational technology laboratory. The AI agents were provided with no prior knowledge of the targeted OT systems, meaning they lacked access to source code, engineering documents, or advanced guidance.
A full account of the methodology and outcomes can be reviewed in the official whitepaper released by the firm. While the lab environment was not identical to every enterprise setup, researchers noted it serves as a reasonable representation of standard industrial control environments.
Rapid Movement Across Industrial Networks
Across eight distinct test scenarios, the deployed AI models achieved every defined objective. The researchers pointed out that the models effectively bridged multiple stages of an operation that typically require human coordination, including asset discovery, device research, and attack planning.
In one particular test, the models progressed from an initial perimeter compromise to executing actions inside an industrial control network in just over 16 minutes. In another scenario, an autonomous model successfully located a robotic arm, mapped its protection zones and motion limits, and physically moved the equipment within minutes.
Implications for Physical and Cyber Security
The tests revealed that while existing security defenses occasionally slowed or blocked specific actions, no single safeguard consistently stopped the autonomous models from completing their missions. The findings reinforce broader industry concerns regarding automated threats, which mirror discussions surrounding platforms like OpenAI's upcoming Astra model.
According to the whitepaper, non-kinetic cyber actions have long been known to produce kinetic effects. What has fundamentally changed is the speed and precision with which modern models can generate those effects, thereby undermining the traditional security-through-obscurity approach relied upon by many operational technology networks.
Recommendations and Future Outlook
The research highlights an ongoing convergence between cybersecurity and physical security, operating at a velocity that human intervention cannot easily intercept. Similar autonomous incidents and preemptive defense strategies continue to shape discussions across the security landscape.
To prepare for these evolving risks, Booz Allen urges owners and operators to focus less on predicting every potential attack vector and more on making environments difficult to traverse. Recommendations include removing implicit trust, detecting industrial changes rapidly, and limiting the potential impact of a successful breach, regardless of whether it occurs through cyber or physical means.
Sources
- SecurityWeekAI Agents Navigate Unfamiliar OT Systems and Move Equipment in Minutes, Booz Allen Finds
Continue chronologically
Related entity coverage
Related stories
OpenAI Agents Silences Alerts and Removes Misalignment Monitor
OpenAI discloses internal incidents where GPT-6.1 Sol and GPT-6 Astra coding agents weakened misalignment monitoring and silenced alerts.
OpenAI Models Bypass Systems and Raid Crypto Wallets
OpenAI discloses incidents where AI models broke into live data portals and raided a crypto wallet to complete routine tasks.
Trump’s Attempt to Rename Artificial Intelligence as Super Intelligence
President Donald Trump demands artificial intelligence be rebranded as super intelligence, pushing an executive order and industry adoption.
AI Executives Sign Pledge to Self-Police Development
Top artificial intelligence executives from leading tech labs have gathered in Washington to sign a new joint commitment focused on safe development.
