OpenAI to Enhance Safety Processes for Paid Tool Customers

OpenAI said it will enhance safety processes for paying users who have access to its most advanced artificial intelligence models, stepping up its safeguards at a time when customers are entrusting AI with more complex tasks and sensitive information.

OpenAI currently pledges “zero data detention” to some customers using its application programming interface, or API, meaning it doesn’t retain prompts or model responses after a request is processed. It also does not have access to customer content. On Wednesday, the company said it has begun testing Private Safety Processing, an enhancement that aims to deter both bad human actors and misaligned AI agents from hacking attempts.

OpenAI is testing with “early customers” such as Microsoft Corp. and Databricks Inc. It expects to roll out Private Safety Processing, and a technical white paper about it, in September.

“We’re seeing with more capable frontier models that often, risks are emerging not just by looking at one single prompt and response pair, but when you look over time at multiple interactions,” said Aleah Houze, OpenAI’s head of product policy, at a press briefing ahead of the announcement.

The development comes as AI companies are increasingly offering so-called agentic tools that can take on complicated tasks requiring multiple steps, without explicit instruction by users.

“An example of this would be somebody might be asking about the weakness in a company’s software in one conversation, and then later in another conversation they might ask about remote access or what security tools can detect,” said Houze. “If you look at just those isolated prompt and response pairs, they might look like very legitimate cybersecurity research in isolation. But when you look at them together in the broader context, you might detect that somebody is actually attempting a cyber attack.”

OpenAI said in a statement that under the new process, customer content remains restricted from the company to preserve privacy.