OpenAI says it can police misuse without keeping your company's data — and it is pointing at Anthropic while saying so

OpenAI previewed a system called Private Safety Processing on 19 August, aimed at enterprise and API customers rather than people paying for ChatGPT. Those customers can already ask for zero data retention, where OpenAI stores nothing. The gap OpenAI acknowledges is that its existing safeguards judge each request on its own, so abuse spread thinly across many conversations is hard to spot. The new system runs automated checks over inputs and outputs across multiple conversations without staff seeing them, and if something trips, OpenAI receives what it calls a narrowly defined signal rather than the prompts themselves. It then decides whether to act and may ask the customer for context, and the customer can volunteer more. Data can stay on infrastructure the customer controls, or sit with OpenAI under encryption keys the customer holds. OpenAI says it will start rolling this out and publish a technical paper in September. The comparison being drawn is with Anthropic, which retains data for 30 days on what it calls covered models for safety purposes.

✓ Verified · 2 sources

Related stories