OpenAI announced new safety measures for paying users of its most advanced AI models. The company wants to protect customers who use the platform for complex and sensitive tasks.
Currently, OpenAI offers a “zero data retention” policy for API users. This means it does not store prompts or model responses after a request is finished, and it does not access customer content.
The new feature, called Private Safety Processing, is designed to spot risky patterns across several interactions instead of just one. It can detect dangerous behaviour that may arise over time.
Private Safety Processing is being tested with early customers such as Microsoft Corp. and Databricks Inc. The company plans to release the tool and a technical white paper in September.
Aleah Houze, OpenAI’s head of product policy, said at a press briefing that “risks often emerge when you look at multiple interactions over time.”
The update comes as AI tools can now perform multi‑step tasks without explicit user instructions. For example, a user might first ask about a company’s software weaknesses and later ask for ways to gain remote access. These separate requests look harmless on their own, but together they could indicate a cyber attack.
OpenAI stated that customer content will remain private under the new process, preserving user privacy while improving safety.