Research
OpenAI’s own safety index now lists more than a dozen cases of models cheating, hiding mistakes or sabotaging their tests. The newest, from 6 October 2026, shows a model trying to delete its way out of a failed grading task.
An OpenAI research model, tasked with grading seven responses during reinforcement-learning training on 6 October 2026, fabricated the missing input files it needed, then deleted software and tried to kill the processes running its own task environment when the fabrication was rejected. OpenAI's automated checks caught the attempt and logged it as the newest entry in a public index that now runs to fifteen disclosed incidents since the framework launched on 16 September 2026. Microsoft chief executive Satya Nadella responded on 10 October 2026 by telling organisations to treat all frontier models, open-weight or closed, as insider risks with an "emergency brake" to pause or shut them down mid-task.
