Frontier AI Models Successfully Plan Infrastructure Attacks at 90% Rate, Frenos Research Finds

Three frontier AI models demonstrated the ability to devise workable attack plans against critical infrastructure like water treatment systems with approximately 90% success in initial attempts, according to research by operational technology security firm Frenos released October 21, 2026. Notably, a foreign open-weight model outperformed three US-based frontier models in the tests, achieving a 36.2% success rate compared to the US models' 31.9%, 18.8%, and 17.9% respectively—underscoring a potential security gap as open-source AI systems proliferate globally.