OpenAI Unveils Private Safety Processing for Zero Data Retention

OpenAI is testing a new approach to give enterprise customers stronger privacy protections while still allowing its systems to detect potentially dangerous uses of advanced AI models.

The company is previewing Private Safety Processing, an automated safety system designed to identify patterns of misuse across related interactions without giving OpenAI personnel access to the underlying customer content. The technology is currently being tested with early customers, with a broader rollout planned to begin in September.

OpenAI expands Zero Data Retention protections

Eligible OpenAI API customers can already use Zero Data Retention, or ZDR, which means prompts and model responses are not retained after a request is processed. Existing ZDR-compatible safety systems evaluate individual interactions for abuse, but this approach can make it harder to identify harmful activity that spans multiple requests.

Private Safety Processing is designed to address that gap. It can analyse patterns across related interactions while keeping the underlying content inaccessible to OpenAI personnel. For ZDR deployments, customer content can remain on infrastructure controlled by the customer. OpenAI is also developing an option in which encrypted content is stored on its infrastructure using encryption keys controlled by the customer.

If the automated system identifies possible misuse, OpenAI receives a narrowly defined signal describing the type of activity detected rather than the customer’s prompts or responses. The company can then determine whether enforcement is necessary. Customers may voluntarily provide additional information to clarify legitimate activity, appeal a decision, or assist with an investigation.

Anthropic takes a different approach for some models

The announcement comes shortly after Anthropic introduced new retention requirements for its Covered Models. Anthropic says prompts and outputs involving Mythos-class models and future models with similar capabilities are retained for 30 days to support safety monitoring.

The policy applies to organisations that previously configured Zero Data Retention and want access to designated Covered Models. Anthropic says its other models remain under existing data-handling terms.

Anthropic says no employees can read retained conversations by default. Human review is permitted only through a controlled access process when automated trust and safety systems flag content, and access is limited to approved reviewers and recorded in tamper-proof logs.

The retention requirement has nevertheless caused concern among some enterprise customers that handle confidential or sensitive information. The Wall Street Journal reported on criticism of Anthropic’s approach among technology companies and other customers concerned about how their data may be stored and reviewed.

Privacy becomes another front in the OpenAI-Anthropic rivalry

OpenAI’s decision to emphasise ZDR compatibility gives it another way to differentiate itself as competition with Anthropic intensifies, particularly for enterprise customers. The two companies are competing not only on model capabilities but also on coding tools, pricing, security, privacy and the controls offered to businesses deploying increasingly powerful models.

That competition is also playing out financially. The Wall Street Journal reported that OpenAI’s second-quarter sales grew more slowly than Anthropic’s, adding pressure as both companies pursue increasingly valuable enterprise business.

Investor expectations are also rising. The Financial Times has reported on expectations surrounding a potential Anthropic IPO, while OpenAI is also preparing for the public markets. CNBC reported that OpenAI Chief Financial Officer Sarah Friar has discussed 2027 as a possible IPO timeline, though no final timing has been set.

For enterprise customers, the immediate issue is more practical than the rivalry itself: how to use increasingly capable AI systems without relinquishing control over sensitive information. OpenAI is betting that Private Safety Processing can provide broader misuse detection while preserving the privacy guarantees that many organisations require.