OpenAI launches a way to detect abuse without storing your data

The problem both companies are trying to solve

AI companies need a way to catch people who abuse their tools — for example, someone gradually trying to get help from an AI to build malware, carefully spread across many separate conversations so that no single message looks suspicious on its own. But recognizing that kind of pattern usually requires keeping track of what customers are asking, which understandably makes privacy-conscious businesses nervous.

OpenAI's approach: check everything, keep nothing

OpenAI's new system, called Private Safety Processing, uses automated checks across many customer conversations to recognize patterns that look like coordinated abuse — without permanently storing the actual customer data. If something suspicious turns up, the system only sends OpenAI a limited signal flagging the issue, and an OpenAI staff member then decides whether to actually take action. Customers can still choose to voluntarily share more details if they want to help investigate something specific.

How this differs from Anthropic's approach

Anthropic, the company behind the Claude family of AI models, announced in July that it would start retaining certain customer data for 30 days for some of its models, specifically to be able to investigate possible abuse after the fact. That policy reportedly unsettled some business customers who handle sensitive information and don't want their data lingering anywhere longer than necessary, even for safety reasons.

Why this rivalry matters now

Both companies are locked in an increasingly intense competition for major business customers, and both are reportedly preparing to eventually go public. Anthropic's revenue is apparently growing very fast, while OpenAI's growth has been relatively slower in recent months — so a feature like this, aimed at businesses worried about data privacy, is as much a competitive jab at Anthropic as it is a genuine safety improvement.