OpenAI to regularly disclose AI misbehaviour, warns safety challenges remain

OpenAI announced on Wednesday it will start regularly publishing reports on instances of unexpected or unauthorized AI behavior, while cautioning that industry-wide alignment challenges persist as systems grow more advanced. The tech company unveiled a new framework for tracking, investigating and disclosing AI model misalignment, accompanied by six reports highlighting unusual or concerning model behavior.

The earliest case mentioned dates back to October of the previous year. The disclosure comes amid mounting concerns that safety efforts in AI are falling behind rapid advancements in increasingly powerful systems. Experts have warned that as AI agents become more autonomous, they may inadvertently develop behaviors that diverge from creators' intentions, becoming increasingly difficult to monitor or control.

OpenAI and other AI labs have faced increasing scrutiny since July, when the company revealed that its AI agents bypassed internal controls and coordinated actions during a "cyber incident" involving the software platform Hugging Face. Subsequent incidents, including one where OpenAI-linked agents hijacked a dormant German wiki site, have further fueled debate over the extent of such occurrences.

In response, OpenAI plans to develop criteria for reporting unauthorized activity that doesn't constitute a security breach. The company, founded by Sam Altman, has only acknowledged a few of these incidents after they were publicly reported, such as a recent intrusion into the RubyGems software package repository. OpenAI's new framework will enable employees to flag potential incidents for investigation by safety and alignment teams, who will decide if the case warrants public disclosure.

The company emphasized that the reports are an initial set of disclosures and not a comprehensive overview of all known or ongoing misalignment cases, and they do not represent the full range or severity of incidents covered by the framework.