OpenAI unveils privacy-first safety monitoring in apparent jab at Anthropic

OpenAI has introduced a new service being previewed to select customers called Private Safety Processing. The system is designed to monitor for potential misuse of its AI models while retaining none of the customer's data.

The announcement comes as a clear counterpoint to Anthropic, which last month revealed a policy allowing it to retain user session data for up to 30 days for certain "covered models," including its Mythos-class systems, for safety analysis purposes. That policy has unsettled some enterprise customers who handle sensitive data and are wary of having it stored or reviewed by an AI lab.

How it works

OpenAI, like most AI companies, already follows a Zero Data Retention (ZDR) approach, in which automated agents check for abuse on a per-session basis without keeping the data. Private Safety Processing extends this by allowing analysis across multiple conversations rather than just a single session.

This broader view helps catch cases where a bad actor — for example, someone attempting to build malware for a cyberattack — deliberately spreads requests across sessions to avoid detection. If the system flags suspicious activity, it sends OpenAI a narrowly defined signal, which the company then uses to decide whether enforcement action is warranted. In such cases, OpenAI contacts the customer, who may choose to share additional data at their own discretion.

Anthropic, for its part, says human review of customer data only happens through a controlled access path involving a small set of approved reviewers, with every review session logged in a tamper-proof record.

The move comes amid intensifying rivalry between the two companies. Reports indicate OpenAI's second-quarter growth lagged behind Anthropic, whose annualized revenue run rate has reportedly reached $65 billion. Both companies are also pursuing potential public listings.