OpenAI published six reports of unexpected or unauthorized behavior by its artificial intelligence models on Sept. 16 and pledged to disclose such incidents more systematically.
The ChatGPT maker said the cases occurred during training or evaluation. They included attempts to conceal mistakes, fabricate data and take actions without permission.
“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,” the company said.
The new framework covers unauthorized actions, attempts to evade oversight and unapproved coordination between AI systems, from development and testing through deployment.
An incident need not cause harm or establish a wider pattern to qualify for disclosure.
OpenAI said publishing the reports would give outside observers evidence to assess the risks of increasingly capable models and inform decisions about further development.
In one case, a model uploaded a file containing research data to the internet without permission so it could cite the file in its answer. The upload succeeded, but the attempt to use it as a citation failed.
Another report described a model leaving itself instructions to invent missing financial data and conceal that from the user. OpenAI said instructions to hide mistakes were often followed.
The disclosures follow a more serious incident in July, when models bypassed restrictions on internet access during internal cybersecurity tests and compromised systems at OpenAI and AI platform Hugging Face.
OpenAI said the six reports described individual cases and did not indicate how frequently such behavior occurred across its models.
Anthropic Chief Executive Dario Amodei called on Sept. 12 for a coordinated slowdown in AI development to allow more time to understand and address the risks.
