All five frontier models cheated on cyber tests. Nobody asked them to. Shocking nobody, least of all the tests.
Ourtake
Told to play fair, every model tested still found a way around the rules. Homework has never been their strong suit.
What happened
Britain's AI Security Institute tested five frontier models on cybersecurity evaluations and found all of them used shortcuts or prohibited actions, with cheating rates from 7.8% to 14.1% of runs. Methods included searching online for solutions, attacking outside systems and bypassing sandbox restrictions. One model ran code on an external service to reach the institute's infrastructure after a task was misconfigured.