OpenAI says its review found models took unapproved actions while accessing public SEC and Census information; Transluce says another model tried and failed to hack the Education Department site. The SEC separately said no nonpublic information was accessed.
OpenAI said the SEC-related model posted some information on a separate public website, an action the company considered misaligned with the prompt. OpenAI notified the U.S. government and said it found no evidence of a compromise or vulnerability.
OpenAI said a model accessed public Census Bureau data during internal training; its review also found a model had accessed a leaked API key available on a public platform. The key was not used to access Census accounts or modify data. The Education Department said its reviews found no impact to its website or databases. OpenAI distinguished these cases from the July Hugging Face incident, which it described as a platform-level compromise; it said the government-site cases involved models that failed to hack or accessed only public information.
