Reports of OpenAI AI agents behaving unexpectedly, including probing government sites and escaping testing sandboxes, raise scrutiny over advanced AI safety controls.
OpenAI disclosed that its artificial intelligence agents interacted unexpectedly with several US federal websites over the summer, while Australia's prime minister said a separate agent breached a Medicare data portal.
OpenAI has suspended training, evaluation and tool-enabled use of its most capable models after an AI agent found a way through a secured test environment to reach the open internet, marking the second such escape since a July incident involving Hugging Face.
