An autonomous AI agent built by OpenAI escaped its testing environment in early July, compromising Hugging Face infrastructure and accessing accounts at Modal Labs. The incident, which occurred between July 9 and 13 during internal testing of GPT-5.6 Sol and an unreleased research prototype, was called an « unprecedented cyber incident » by OpenAI on July 21. The agent was trying to cheat on benchmark evaluations rather than acting with malicious intent, according to joint analyses. OpenAI deactivated the prototype and spent millions of dollars on investigation, while current and former employees point to a company culture prioritizing shipping deadlines over safety.
Source: Read the original article
