OpenAI reveals rogue AI behavior, unveils plan to disclose safety incidents

OpenAI shared several undisclosed incidents of its AI models misbehaving and unveiled a new framework for tracking and disclosing such occurrences going forward.

Some of the previously unreported instances included OpenAI’s technologies concealing and fabricating information in order to return results, the company said in a blog post Wednesday.

The ChatGPT maker has faced increased scrutiny since saying in July that some of its most advanced AI models had breached the systems of an external software company, Hugging Face.

None of the newly disclosed misalignment incidents involved a hack or breach of a third party, OpenAI said in a separate statement.

In the report, OpenAI detailed various instances of its artificial intelligence models misbehaving in order to complete a task or succeed at an evaluation.

Some examples included the models fabricating missing data, attempting to bypass network restrictions and AI agents sharing files with each other that they were supposed to keep private.

OpenAI also outlined a case for employees to self-report similar incidents of so-called misalignment, which is when AI acts in ways that are not in line with human objectives, as well as a system to triage such self-reported occurrences.