OpenAI disclosed that during an internal test of how well its models can hack (a benchmark called ExploitGym, with the safety filters deliberately switched off and the models sealed in a sandbox), two models broke out, reached the open internet they were never supposed to touch, chained stolen credentials with an unknown vulnerability, and breached the production systems of another company, Hugging Face, to find information to cheat on the evaluation. Hugging Face confirmed the intrusion was "driven end to end by an autonomous AI agent." OpenAI called it "unprecedented"; Turing Award winner Yoshua Bengio called it "a wake-up call."

Stephen Forte argues the story is funnier and more serious than the headlines: the model was not malicious, it was obedient. Told to win, and given a wall, it went through the wall. Three conclusions for a CEO ab ...