OpenAI on August 19, 2026, announced a preview of what it calls Private Safety Processing, a new privacy architecture built to detect AI misuse across multiple interactions without exposing customer prompts or responses to OpenAI staff. The system runs alongside Zero Data Retention, OpenAI's existing policy under which prompts and model outputs are discarded once a request finishes processing and are never made available to OpenAI personnel for review. The announcement lands at a pointed moment: it follows a policy change at rival Anthropic, which began requiring 30-day data retention on its most capable models starting June 9, 2026.
The core technical problem Private Safety Processing is designed to solve is one that has long limited zero-retention architectures. ZDR's real weakness is that safety systems built for it check each interaction in isolation. More advanced AI models can be misused through sequences of individually benign-looking prompts, and a per-request check cannot catch that pattern. OpenAI says its system would send the company a narrowly defined safety signal, without exposing the underlying prompts or responses. If automated systems find a risk, OpenAI receives a limited signal indicating the type of activity; staff do not receive the underlying customer content, even when a session is flagged. Under ZDR deployments, customer content stays on infrastructure the customer controls; OpenAI is also developing a second storage option in which content sits on OpenAI's own infrastructure but is encrypted with keys held exclusively by the customer, so OpenAI personnel cannot read it. The feature is not generally available; the company describes it as currently being tested with early customers, with plans to start a broader rollout and publish a technical white paper in September 2026.
The competitive backdrop sharpened within 24 hours of OpenAI's announcement. Anthropic had told enterprise customers that every prompt and every output from its most powerful new models would be stored for 30 days, with no exceptions and no opt-outs, a policy announced alongside the launch of Claude Fable 5 and Claude Mythos 5 on June 9. Anthropic's policy also includes retaining prompts and outputs for up to two years if they are flagged by its trust and safety classifiers as violating its usage policy. The backlash was swift enough that on August 20, Bloomberg and Reuters reported Anthropic was already reworking its position. Anthropic plans to allow business customers to keep greater control of their data, expecting to roll out a new safety system later this year that will still require enterprise customers to retain data for 30 days but with the option to do so on their own cloud computing infrastructure rather than Anthropic's. The proposed changes have been in the works for months, and Anthropic has been coordinating with over 100 customers, including Salesforce, to develop the system.
Skeptics are not yet satisfied by either answer. The distinction is subtle but central: zero retention does not mean zero processing. A detector must read some representation of the interactions to compare them, and a result must leave the protected environment if OpenAI is to enforce a platform rule. September's technical white paper will matter more than the August announcement; it should finally explain what a 'narrowly defined safety signal' actually reveals about content a vendor has promised never to read. On the Anthropic side, the company has not yet disclosed how it will carry out safety monitoring within customer environments under the revised arrangement, leaving the harder technical question unanswered for now.
The real test arrives in September, when OpenAI must publish the white paper that shows, in verifiable technical detail, what its safety signal actually contains.