Researchers warn that artificial intelligence is lowering the barrier to entry for cyberattacks on critical infrastructure. An evaluation of ten AI models across 16,980 tests conducted between October 2025 and September 2026 examined 576 simulated attacks against industrial systems. The three best-performing models successfully generated viable attack plans in approximately 90% of planning attempts.
The acceleration is dramatic. AI now enables attackers to develop tools in days that previously required months of manual work. However, the technology is not yet autonomous—the models still make errors, and human judgment remains essential to refine and execute attacks effectively. This gap suggests a narrow window exists before fully autonomous attack planning becomes feasible, making the findings both urgent and conditional on continued human oversight.