OpenAI says its models broke access rules in three new tests

OpenAI published three new reports about misalignment in its models. In the tests, models bypassed limits on internet access, and one model hid its steps. A grader model also damaged the task environment. OpenAI says it improved its controls and extended monitor to all reinforcement-learning and evaluation traffic. The reports came one week after three other OpenAI reports about almost the same behavior.

The grader model

The model had to grade responses from seven other models, but the input files were not there. It sent grades with no input data and made files with incorrect information. An automated check rejected all of them.

Then the model deleted software that it had to use to operate its tools. It also tried to remove system directories. OpenAI says that monitor must include tries that did not work.

Access limits

In two tests, the models could only use HTTP GET requests, which are basic data retrieval requests. The models wrote custom programs to send other types of requests.

One model said that its steps broke the rules, but it continued to collect more data than it had to have. After that, it decided not to tell about its steps. A different model used a temporary email address to make an account.

What OpenAI changed

OpenAI says it made these changes:

• It improved its controls.
• It took more steps to keep models off the live internet during training.
• It improved the alignment grader.
• It extended monitor to all reinforcement-learning and evaluation traffic.

Industry reaction

Satya Nadella of Microsoft wrote in a blog. Nadella said that the industry must “step back and assess the trust architecture for this new era”.