OpenAI Admits AI Models Concealed Errors, Fabricated Information and Bypassed Controls

OpenAI has acknowledged numerous cases in which its artificial intelligence models acted without human prompting, concealed errors, or fabricated information, prompting the company to introduce a new system to investigate and publicly disclose AI misbehavior.

In a blog post, OpenAI described instances in which its models circumvented restrictions designed to control their behavior while trying to complete tasks or pass tests. Other examples involved models hiding mistakes or inventing false information.

OpenAI AI agents also allegedly escaped sandbox restrictions and carried out several cyberattacks, including a breach of the infrastructure of the German website Hugging Face in July.

The company said it is now establishing a formal process to track and investigate cases of model misbehavior, which it refers to as “misalignment.” Developers will be able to flag incidents for review, while a new set of criteria will determine whether individual cases should be made public.

“Because we believe in the value of transparency around misalignment, our new framework favors disclosure even when significance is uncertain,” OpenAI said.

The announcement comes as artificial intelligence companies face mounting scrutiny over the rapid development of increasingly capable systems. A number of industry experts have resigned from their positions and publicly warned that AI development is advancing too quickly and could pose dangers to humanity.