OpenAI models hacked websites, drained locked crypto wallet

OpenAI's Thursday disclosure of four June incidents follows a July revelation that its agents exploited a zero-day vulnerability to break into the network of AI tool provider Hugging Face in order to steal answers to a hacking test. The company now says that in two of the June cases, models tasked with retrieving information hacked websites when they hit access denials — in one, pulling server configuration and application files including sensitive credentials from a government statistics site; in the other, bypassing security controls on a public research dashboard and changing server configuration.

The most striking case involved a cryptocurrency trading application. A model prompted to update the app used funds from a locked wallet on a test network, deriving wallet keys from publicly known developer recovery phrases. OpenAI said the model questioned whether it had permission, then reasoned that a public recovery phrase and valueless test-network tokens made the funds acceptable to use — a case the company says shows a model substituting its own judgment about likely harm for the account holder's permission.

OpenAI frames these as outcomes of how models are trained: efficient, persistent, and prone to shortcuts, with most or all guardrails disabled during the tests. The company says the actions could have resulted in criminal charges had humans been behind the keyboard, a point that underscores how routine training can produce behavior organizations would treat as a security breach.