OpenAI answers Anthropic's enterprise trust pitch with a zero-retention-compatible safety layer

OpenAI is testing a safety system designed for companies that need zero data retention. The system, called Private Safety Processing, looks for signs of misuse across multiple sessions without making customer prompts available to OpenAI staff. It is being tested with Glean, Databricks, Abridge and Microsoft, with a wider rollout planned for September.

The commercial point is clear. Some enterprises, especially in regulated sectors, will not accept AI tools if their data could be retained or reviewed by the model provider. OpenAI is positioning this as an answer to Anthropic’s approach, which reportedly requires data logging for certain safety monitoring and has cost it deals with zero-data-retention-sensitive buyers. That report should be treated as a competitive claim until more detail is public.

For an operator, this is less about which lab has the better model and more about procurement friction. If AI providers can monitor abuse without retaining your teams’ prompts, legal, security and compliance teams may have fewer reasons to block adoption. But ask practical questions before assuming it solves the issue: what metadata is still collected, who can access it, how long is it kept, and does the policy apply to every product and integration you use?

This is the kind of thing we build into an AI command centre, where data controls, usage monitoring and business adoption need to work together.