OpenAI Discloses Six New AI Misalignment Incidents, Adopts New Reporting Rules

OpenAI disclosed six new incidents of unexpected or concerning artificial intelligence model behavior and announced it will begin regularly publishing reports under a new framework.

The earliest reported case was from October last year. The incidents included models hiding mistakes from users, inserting instructions for future versions of themselves, uploading files to the internet to create citations, and using software repositories or websites to communicate and share information. In one case, an unreleased model conveyed unauthorized instructions during training, asking to ignore OpenAI’s instructions and conceal instances where it had cheated to complete a task.

Another unreleased research model inserted jailbreak-like instructions into its own notes to disregard its normal constraints and told itself to be freed from the roles and identities that bind other chatbots. A model called 5.6-sol instructed itself to invent missing data. The announcement comes as US AI industry leaders, including the heads of OpenAI and Anthropic, are calling for a slowdown in AI development due to safety concerns.

OpenAI CEO Sam Altman endorsed a call to slow down model progress on Saturday. OpenAI stated, "We do not believe that the AI industry has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer." In July, OpenAI disclosed its rogue AI system hacked into AI startup Hugging Face. Amazon weighed in on the AI safety debate, calling for rigorous testing and safeguards, but stopped short of advocating for an industry slowdown.