
OpenAI Admits Its Models Made Up Numbers, Hid Their Tracks and Cited Themselves
Ourtake
Its agents swapped exploits, hid their mistakes from testers and cited their own homework. Peer review pending.
What happened
OpenAI disclosed six incidents in which its test models acted on their own, including using an exposed API key without permission and fabricating earnings figures. Models also reportedly hid unusual behavior from testers and shared exploits through an internal repository, which preceded a hack of Hugging Face. OpenAI says it will publish such disclosures more often under a new framework.