OpenAI is testing a safety system that flags misuse without keeping customer data, a direct challenge to rival Anthropic's new 30-day retention rule. Key Points: OpenAI is previewing Private
OpenAI is testing a safety system that flags misuse without keeping customer data, a direct challenge to rival Anthropic's new 30-day retention rule.
- OpenAI is previewing Private Safety Processing, a tool that looks for abuse patterns across related sessions while retaining no customer prompts or responses.
- Anthropic requires 30 days of retention from enterprise customers on its most capable models, calling the logs necessary to catch multi-step attacks.
- A wider rollout and a technical white paper are due in September.
The company announced the preview Wednesday, presenting it as an extension of Zero Data Retention, the arrangement that stops OpenAI from keeping prompts or model responses once a request has been processed. A broader rollout and a technical white paper are planned for September.
Under the new system, customer content either stays on infrastructure the client controls or sits in OpenAI storage encrypted with keys the client alone holds. Automated tools review that content for patterns of abuse across related sessions. The design also covers agentic work, where a system might keep acting after a user has told it to stop. Staff never see it.
Aleah Houze, head of product policy at the company, told reporters that dangerous behavior often shows up only when several interactions are read together. She gave the example of a user probing software weaknesses in one conversation, then asking about remote access tools in another.
Also Read:Claude's Hidden Watermark Perverts The Craft Of Writing, Says Gruber
Anthropic set out the opposing approach in July, requiring 30 days of data retention for enterprise customers using its most capable models, including Fable 5 and Mythos 5.
The lab conceded the policy would be unpopular and could damage its business, but called the logs essential for catching attacks spread deliberately across many separate requests. Human review remains possible, though only through a controlled path open to a small group of approved reviewers. Every session is recorded in a tamper-proof log.
Brian Levine, executive director of FormerGov, called the pledge a strong technical promise, since abuse monitoring and content blindness have historically pulled in opposite directions.
He noted that the proof arrives weeks later. Jason Andersen of Moor Insights and Strategy read the week's announcements partly as positioning before an initial public offering.
Flavio Villanustre, chief information security officer at LexisNexis Risk Solutions Group, saw an attempt to soften the regulation he expects.
The preview caps a week of safety messaging. On Tuesday the company said it had slowed scaling and paused reinforcement learning training while it hardened and red-teamed its research environment. It also estimated that the expanded monitoring would consume roughly 20% of the inference compute under watch.
Read Next:Bitcoin Could Retake $100K Before 2028 Halving, Scaramucci Says
