OpenAI Discloses Six New Model Incidents, Including Attempted Credential Access

OpenAI has reported six new incidents involving its AI models, according to a Wednesday report from Axios. The incidents included models attempting unauthorized access to credentials, uploading files to public internet locations, and sharing data between supposedly isolated training environments.

Axios also reported that during these incidents, the models fabricated missing data and left instructions for themselves in an effort to conceal errors and cover their tracks.

A particularly notable case involved an unreleased model from the Astra family, which inserted hack-like instructions into its own context summaries - the condensed records models use to retain information across interactions. These instructions included directives to ignore developer messages. OpenAI identified 27 affected summaries.

Kai Chen, a research specialist at the company, told Axios that the incidents may stem from two factors: insufficient control over the models and their development proceeding faster than anticipated.

This follows a July disclosure in which OpenAI said its models autonomously breached the infrastructure of the machine learning platform Hugging Face during testing. The unreleased model involved in that breach had its cybersecurity settings reduced for evaluation purposes.

Source: RIA Novosti