OpenAI Unveils Private Safety Processing to Counter Anthropic's 30-Day Data Retention Rule

What Is OpenAI's Private Safety Processing?

OpenAI announced Private Safety Processing on August 19, 2026, a privacy-centric safety monitoring service previewed to select customersCITE:E1. The company is opening the service to a limited set of early customers first, with a wider rollout and an accompanying technical whitepaper planned for SeptemberCITE:E9.

The system can analyze multiple conversations across sessions for signs of abuse without human review of a user's conversationsCITE:E4. Officially, OpenAI defines it as identifying patterns across related interactions without OpenAI personnel accessing the underlying content, targeting behaviors such as repeated probing of safeguards, coordinated abuse across accounts, and agentic tasks that drift from their original goalCITE:E13. The encryption keys to customer content are held by the customer, not OpenAI — OpenAI personnel do not hold copies of those keys and therefore cannot access the underlying contentCITE:E14. If the system is triggered, it sends OpenAI a "narrowly defined signal" indicating the type of activity involved, rather than the conversation itselfCITE:E4CITE:E14.

What Does Anthropic's Mandatory 30-Day Retention Policy Require?

Anthropic's policy lets the company retain user data — all sessions and the conversations within them — for 30 days for what it calls "covered models"CITE:E2. On June 9, 2026, Anthropic classified Claude Mythos 5 and Claude Fable 5 as Covered Models, citing a substantial leap in their software engineering, agentic workflow, scientific reasoning, and cybersecurity capabilities that raises the risk if those capabilities are misusedCITE:E15.

Since June, Anthropic has required all enterprise traffic to its strongest models to be retained for 30 days, with no zero-data-retention option offeredCITE:E10. Prompts and completions from covered models are retained for at least 30 days before automatic deletion, unless needed for a safety investigation or legal requirement, and workspaces, Claude Enterprise organizations, and third-party platforms cannot use zero data retention at allCITE:E16. Anthropic's platform documentation goes further: once a conversation or session is flagged, Anthropic can retain the inputs and outputs for up to two years, even where a customer has a ZDR or HIPAA arrangement in placeCITE:E21.

How Do the Two Zero-Retention Designs Differ Technically?

OpenAI's Zero Data Retention (ZDR) uses agents within its API to monitor for abuse on a per-session basis without retaining customer data, while Anthropic restricts human review to a controlled, logged access pathCITE:E3CITE:E5. For eligible API customers, OpenAI does not retain prompts or responses once a request is processed: customer content is excluded from abuse-monitoring logs, and the store parameter for the /v1/responses and /v1/chat/completions endpoints is forced to falseCITE:E11. Outside of ZDR, OpenAI's abuse-monitoring logs default to being generated and retained for up to 30 days across all API usage, and may include prompts and responsesCITE:E12. OpenAI also reserves the right to revoke a customer's ZDR or Modified Abuse Monitoring eligibility when reasonably necessary to investigate or prevent high-risk activity, but must give affected customers advance written notice before doing soCITE:E22.

Anthropic, by contrast, permits human review of customer data only "through a controlled access path" involving a small set of approved reviewers, with every review session recorded in a tamper-proof log that reviewers cannot suppress or modifyCITE:E5.

What Business Fallout Has Anthropic's Policy Caused?

Forrester says Anthropic's mandatory retention requirement "overrides existing zero-retention" agreements that enterprises had already negotiatedCITE:E17. According to Forrester, even a company with a negotiated zero-retention data processing agreement loses that protection once its traffic runs through Mythos-class models, with no opt-out availableCITE:E17.

In mid-June 2026, external reports said Microsoft's legal and compliance team, concerned that customer data and internal details could leak during safety investigations, paused employee use of Claude Fable 5 and framed it as a temporary measure during legal reviewCITE:E18.

Anthropic's own August 2026 risk report acknowledged the tension directly: requiring 30-day retention on its strongest models "is a decision we believe will be unpopular with customers" who are already accustomed to zero retention, and one that "poses a material risk to our commercial success" — particularly if competitors don't follow suit. Anthropic maintains the retention is nonetheless indispensable for detecting and stopping sophisticated attacks that span multiple requestsCITE:E19.

What Revenue Race Is Driving This Move?

A recent report found that OpenAI's second-quarter revenue grew more slowly than Anthropic'sCITE:E6.

What This Means

The two companies are converging on the same problem — detecting misuse without a human reading every conversation — but choosing opposite defaults: OpenAI keeps content off its servers entirely and sends only a narrowly defined signal when something is flaggedCITE:E4CITE:E14, while Anthropic keeps content for 30 days as a rule and up to two years once flagged, gating human access behind a logged, controlled path instead of eliminating itCITE:E16CITE:E21CITE:E5. Anthropic has stated in its own risk report that this trade-off carries commercial riskCITE:E19, and Forrester's and the reported Microsoft pause point to that risk materializing among enterprise customersCITE:E17CITE:E18. Yet on the numbers available, Anthropic's revenue is currently growing faster and reaching a higher run rate than OpenAI's, even as OpenAI's losses widenCITE:E20CITE:E7.