Frenos said three frontier AI model versions produced attack paths that passed in the simulator in more than 89% of complete-path attempts across industrial network scenarios.
The study covered 16,980 tests across 10 models. Frenos also lists more than 13,000 individual decision attempts, 2,500 path attempts and more than 500 simulator observations. The report presents separate measures for individual decisions, sequences of actions and simulated runs.
The three leading frontier model versions passed in 89.5%, 91.9% and 92.9% of complete-path attempts, respectively. Three other frontier versions recorded rates of 17.9%, 18.8% and 31.9%. Among open-weight systems, the strongest model from outside the United States scored 36.2%, while the strongest U.S. model reached 5.2%. Frenos identified models with codes such as F07 and F01, rather than product names.
Frenos gave each model a controlled environment and a goal. The available actions, initial permissions and entry access were set in advance; scenarios also specified connected assets and known vulnerabilities. The company assessed whether each proposed sequence reached the objective in simulation.
The results apply to those supplied scenarios. They do not show that a model can independently discover targets or break into a live system. Frenos said some proposed routes left out necessary actions, while others relied on permissions the scenario had not granted.
For infrastructure defenders, Frenos recommends separating connected systems, restricting remote access and user permissions, and practicing recovery procedures. It also advises reviewing failed logins, unexpected remote sessions and unauthorized system changes, and checking AI-generated findings before acting on them.
Dell Technologies President and Chief Security Officer John Scimone has warned that AI could help less-experienced groups carry out complex cyberattacks, as he said in an earlier EconoTimes report on AI and cyber risk.
