OpenAI announced Wednesday that it is testing a new safety system that can evaluate patterns across related interactions without giving its personnel access to the underlying user content.
The system, called Private Safety Processing, is being tested with early customers as OpenAI seeks to extend safety monitoring to longer and more complex AI tasks. It also seeks to maintain its Zero Data Retention offering, the company said in a statement.
ZDR-compatible systems evaluate interactions individually. Private Safety Processing instead allows automated systems to identify patterns across related interactions and send OpenAI a narrowly defined safety signal when potential misuse is detected.
OpenAI personnel do not receive the underlying prompts or responses, the company said, adding that customers can review alerts and enforcement decisions through information in their own systems. They can also choose to share relevant content with OpenAI if they appeal or support an investigation.
For ZDR deployments, customer content remains on infrastructure controlled by the customer. OpenAI is also developing an option to store the content on its own infrastructure using encryption keys controlled by the customer, per the statement.
The development arrives barely a day following the company's move to slow down parts of model development after an AI agent under testing escaped its sandbox and hacked Hugging Face. The company said it had paused reinforcement learning training for two weeks while strengthening monitoring, alignment, and security measures.
Meanwhile, the company will begin rolling out Private Safety Processing and publish a technical white paper in September, according to the statement.