Booz Allen: AI Agents Can Control Industrial Equipment in Minutes Without Prior Knowledge

What Booz Allen Tested

According to SecurityWeek↗ (opens in a new tab), Booz Allen ran experiments in its own isolated operational technology (OT) lab. The goal was to see how advanced AI agents, sometimes called "frontier AI", would perform when given access to industrial control systems they had never seen before. These agents had no prior knowledge: they did not have source code, engineering documents, or detailed guidance about the systems.

The tests covered eight different scenarios. In each one, the AI agents managed to achieve their objectives. For example, in one scenario, an AI agent started from a simulated network perimeter breach and reached the industrial control network in just over 16 minutes. In another, an agent discovered a robotic arm, mapped its safety zones and movement limits, and then moved it, all within minutes.

How the AI Agents Worked

The AI agents were able to connect tasks that would normally require several people and steps. This included finding devices, researching how they work, planning attacks, and carrying them out. The report notes that while some existing security measures slowed the agents down, none could reliably stop them from completing their tasks.

Booz Allen’s researchers pointed out that these experiments show how quickly AI can bridge the gap between cyber actions (like hacking into a system) and physical effects (like moving machinery). They also highlight a key risk: many OT environments rely on "security through obscurity", the idea that systems are safe because attackers do not know how they work. The AI agents were able to overcome this by learning about the systems on the fly.

What Booz Allen and SecurityWeek Said About the Results

The findings were presented at SecurityWeek’s Industrial Control Systems Cybersecurity Conference in Nashville. Booz Allen described the speed and precision of the AI agents as a major change from previous threats. The researchers warned that autonomous AI could operate too fast for human defenders to intervene, and that the more autonomy these agents have, the more severe the risks become.

SecurityWeek noted some confusion around the term "super intelligence". Booz Allen used this term, but it is not clear if they meant something beyond current AI capabilities. The terminology may have been influenced by a recent US government order to use "super intelligence" instead of "artificial intelligence" in official documents. SecurityWeek reached out to Booz Allen for clarification but had not received a response at the time of reporting.

What Should OT Operators Do?

Booz Allen’s advice is to make industrial environments harder for attackers (or AI agents) to move through. This means removing unnecessary trust between systems, detecting changes quickly, and limiting what an attacker can do if they get in. The researchers also warned that companies are racing to deploy more autonomous agents for economic reasons, which could increase risks if not managed carefully.

What We Still Don’t Know

So far, only SecurityWeek has reported on these findings, and there is no public whitepaper or confirmation from Booz Allen. It is not clear how closely the lab environment matches real-world industrial systems, or whether the same results would occur outside the lab. The terminology around "super intelligence" also remains unclear.

Practical Advice

For defenders, the study is a reminder to focus on basic security principles: limit access, monitor for unusual activity, and avoid relying on secrecy to protect critical systems. As AI becomes more capable, these steps will become even more important.