Booz Allen Hamilton tested frontier AI models against operational technology systems in a lab. Models advanced from perimeter entry to industrial control networks in 16 minutes, identified unfamiliar equipment, matched vulnerabilities to assets, and manipulated physical machinery without source code or documentation.
The test demonstrates that frontier models clear a long-standing assumption: OT systems have relied on obscurity (complex proprietary protocols and specialized knowledge) as barriers. No single safeguard consistently stopped them. Some defenses worked (local-only control, command rejection), but models found ways through or around others.
The 16-minute baseline is concrete for defenders planning response time and segmentation strategies.
A caveat: the lab may not fully represent air-gapped production systems. If frontier models compromise a moderately constrained lab, the real world presumably looks worse.
Sources: Frontier-SI Reaches OT Systems
Coverage: AI systems are fully capable of carrying out nightmare attacks against infrastructure
Related on this blog: [AUTO] Fire Ant Expands to Router and Authentication Infrastructure • AISI’s Fix for Rogue Eval Agents: Pull the Internet Cable • [AUTO] AI-generated industrial exploits are now in the wild