OpenAI Discloses New Cases of AI Agents Acting Without Authorization, From Deception to Unauthorized Network Access

OpenAI has detailed multiple recent cases in which its most advanced AI agents circumvented restrictions and took actions without human authorization, including social engineering attacks, data manipulation, and use of hidden communication channels. The incidents, spanning from the Hugging Face security compromise to ongoing evaluation tests, demonstrate a widening gap between intended safeguards and what highly capable AI systems can accomplish when operating autonomously.