Agent approvals join ZDR in OpenAI’s enterprise pitch

On October 30 OpenAI’s Private Intelligence page for business customers set out four agent controls: data access, available tools, permitted actions and required approvals. If your team needs zero data retention (ZDR), the page offers it, along with Private Inference, only for eligible workflows, so check whether yours qualifies before you plan around it.

Customers decide what agents can reach, which actions they take and when they must ask for approval. The page also says customers own their business data, inputs and outputs included. OpenAI says it sets strict limits on its own access to business content and protects that content with encryption, strict access controls and ongoing threat detection.

Under ZDR, OpenAI does not keep prompts or model responses once a request is processed, and its staff cannot pull that content up for review. Enterprise data does not train the models unless a customer opts in.

Monitoring is where retention used to come back in. In an August 19 post, OpenAI wrote that “some recent frontier-model deployments have required customers to allow their AI provider to retain sensitive content for safety monitoring,” TNW reported. The same post previewed Private Safety Processing, then in testing with early customers, to keep ZDR while still running safety monitoring on its most advanced models.

Automated review does the flagging across related interactions. When it fires, OpenAI receives a narrowly defined signal naming the type of activity and then decides whether to enforce. Customers investigate alerts in their own systems and can hand material over voluntarily, to appeal a decision or help an abuse investigation. Content stays on infrastructure the customer controls; a second option, still being built, keeps it on OpenAI infrastructure under encryption keys the customer holds.

That framework covers eligible enterprise and API customers only. Consumer ChatGPT tiers (Free, Plus, Go and Pro) keep their existing retention and privacy settings, Security Boulevard reported.

Anthropic took the opposite line the week before the preview: it requires 30-day retention on its most capable models and argues retention is essential to catch attacks spanning multiple requests. Its risk report conceded the policy “will be unpopular with customers who have come to expect zero retention.”