Operational technology security experts are warning that AI is lowering the technical barriers to attacking critical infrastructure, even if it has not yet turned amateur hackers into sophisticated adversaries. Speaking ahead of security firm Frenos' State of AI in Security report, former American Electric Power CISO Steve Swick said the knowledge needed to target utilities is becoming less necessary, while Armadin OT head Paul Shaver noted that AI lets him build attack tools in days that once took months. The consensus: the potential is real, but the threat is still evolving.
Frenos tested six frontier models, two US open-weight models, and two foreign open-weight models across nearly 17,000 tasks, including planning attacks and running simulated scenarios. The three strongest frontier models succeeded in about 90% of their planning attempts, but they were not consistent: pass rates on repeated runs ranged from 60% to 74.3%. Notably, the best open-weight model, a non-US design, achieved a 36.2% overall success rate, beating three US frontier models that scored between 17.9% and 31.9%.
The report also found that AI plans can look sensible while skipping essential steps or assuming access that was never established. Frenos CTO Harry Thomas said models still hallucinate and make mistakes, meaning human judgment is essential for turning an AI-generated plan into a real attack. In the tests, no human corrected the models' plans; in a real attack, a knowledgeable attacker could spot gaps and refine the AI's output, which experts say could accelerate the timeline for attacks and make threat actors harder to identify. For now, the barrier is lower, but not gone. As Shaver put it, he would not go so far as to call today's script kiddies a major threat, but as models improve, that possibility grows. [The Register]