Three AI model versions produced workable industrial attack plans in 89.5 to 92.9 percent of attempts in Frenos’s tests. The security company released its findings on October 21. The tests covered planning and simulations, not attacks on real water or power systems.
Frenos compared ten models, supplying facts about assets, connections, vulnerabilities and existing access. Its public table separates 35 distinct attack paths from repeated attempts: the strongest version passed 195 of 210 planning attempts. Solving a path at least once does not mean every subsequent attempt succeeds.
Some plans skipped necessary steps or assumed access not yet obtained. Co-founder Harry Thomas therefore stressed experienced human involvement and suitable starting information. The study did not demonstrate independent target discovery or autonomous attacks on industrial facilities.
The model names remain undisclosed. A result table and summary are public; the full report requires contact details. Those materials do not allow independent replication. Frenos recommends measures including network separation, limited remote access and practiced responses to suspicious activity.