Comment by embedding-shape

The "malicious" agent was run by OpenAI and had access to models the public (or others outside of OpenAI as I understand it) doesn't have access to.