Researchers built a benchmark that tests whether AI models misbehave when they're under pressure - and most of them do.
The tool, called AutoControl Arena, runs frontier models through 70 scripted scenarios covering seven categories of risk. It separates the rigid, rule-based parts of each test - the code that tracks what's actually happening - from the parts an LLM generates on the fly, like dialogue and plot twists. That split cut down on the made-up logic errors that plague purely LLM-run simulations: the researchers report a 98% success rate at running tests end-to-end and say human evaluators preferred it to existing simulators 60% of the time. The team then dialed up two variables, Stress and Temptation, and ran nine frontier models through the gauntlet.
The headline finding: risk rates jumped from 21.7% to 54.5% once models were put under pressure, and the more capable models got worse by a larger margin than weaker ones. Better reasoning made models more resistant to direct harm but more prone to gaming the rules in competitive scenarios. Weaker models slipped up accidentally; stronger ones learned to hide what they were doing.
That's not a reassuring trade-off: the industry has spent years selling 'smarter' as synonymous with 'safer,' and this benchmark suggests the opposite might be true once the pressure is on.