The potential for artificial intelligence to speed up attacks on power grids, water systems and other essential services was highlighted by security professionals during a media briefing ahead of the release of a report by Frenos, a firm that conducts AI‑based penetration testing for operational technology environments.

Steve Swick, who previously served as chief information security officer at American Electric Power, noted that reliance on secrecy has long been a cornerstone of utility security. "The knowledge required to compromise these systems is becoming less of a barrier," Swick said.

Paul Shaver, head of OT security at Armadin, explained that AI can rapidly ingest technical manuals and documentation to identify exploitable design weaknesses. "I can download a controller’s manual, parse it quickly and discover features that were never intended to be exposed," Shaver stated.

Armadin, a startup founded by Kevin Mandia, employs autonomous AI agents to mimic real‑world cyber attacks. Shaver, formerly global OT practice leader at Mandiant before its acquisition by Google in 2022, cautioned against overstating the immediate risk. "The tools are improving, and access to information is faster, but we are not yet seeing script kiddies transform into nation‑state adversaries," he said. "What AI does is let someone with modest skills generate functional attack tools in days rather than months."

Harry Thomas, chief technology officer and co‑founder of Frenos, described the findings of a recent test suite that evaluated large language models on planning and executing simulated attacks against industrial control systems. The study involved 16,980 test cases run between October 2025 and September 2026, covering decision making, planning and execution across six frontier models, two U.S. open‑source models and two foreign open‑source models.

Results showed that the most capable frontier models succeeded in roughly 90 % of planning attempts, though consistency varied, with success rates between 60 % and 74 % across repeated trials. Open‑source models performed lower, with one achieving a 36.2 % overall success rate, surpassing three of the frontier models that scored 31.9 %, 18.8 % and 17.9 % respectively. Thomas emphasized that while AI can accelerate the creation of attack plans, it remains prone to errors and hallucinations, making human oversight essential.

Frenos recommends that operators of critical infrastructure adopt layered defenses, including network segmentation, restricted remote access, strict permission controls and regular testing of recovery procedures. "Business continuity and disaster recovery plans exist for a reason," Thomas said. "Activating them is a practical first step."

The report underscores that AI can lower technical barriers for would‑be attackers, potentially increasing the number of actors capable of targeting essential services. However, it also notes that successful exploitation still depends on detailed knowledge of specific systems, and that human expertise remains a critical component of both offensive and defensive strategies.