Frenos tests show AI attack plans on industrial systems remain uneven
Original title: AI has 'potential' to turn script kiddies into sophisticated infrastructure attackers
Across 16,980 tests, the top three frontier models produced workable plans about 90% of the time, but only 60% to 74.3% passed every repeat.
Frenos’ State of AI in Security report, discussed by OT specialists, tested whether language models can help plan attacks on industrial systems. From October 2025 to September 2026 the company ran 16,980 tests on six unnamed frontier models and four open-weight models, covering action choice, planning, and simulated attacks. The strongest frontier models succeeded on about 90% of planning attempts but were inconsistent across repeats, and a non-US open-weight model led that group at a 36.2% overall success rate. Experts said AI speeds reconnaissance and tooling but does not yet turn low-skill attackers into nation-state operators.
Unconfirmed: Frenos did not name the models or providers, and tests used a simple harness with no human correction.
- Tests included 13,884 decision attempts, 2,520 planning attempts, and 576 simulated attacks.
- Action-choice pass rates ranged from 6% to 47% depending on the model.
- Repeat-pass rates for the top three frontier planners were 60%, 62.9%, and 74.3%.
- Frenos recommends network segmentation, limited remote access and permissions, and tested recovery plans.
- 1.Segment OT networks, limit remote access and permissions, and exercise recovery plans.
- Organizations
- FrenosArmadinAmerican Electric Power
From the original article
4 min read at theregister.comWhen it comes to cyberattacks that shut down power grids, drinking-water systems, and other critical infrastructure, AI doesn’t – yet – turn script kiddies into sophisticated, nation-state attackers. But “the potential is there,” according to operational technology (OT) security experts.
By Jessica Lyons · theregister.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.