OpenAI is rolling out a new privacy-centric system for detecting misuse of its AI models, a move that directly challenges rival Anthropic's recent data-retention policy. The service, called Private Safety Processing, is being previewed to select customers and uses automation to watch for potential abuse without retaining any customer data.

This stands in contrast to Anthropic's July policy, which keeps user sessions and conversations for 30 days on its "covered models," including all Mythos-class models and similar future ones. That policy has frustrated some enterprises handling sensitive data, who worry about their information being stored or inspected.

OpenAI's approach builds on its existing Zero Data Retention (ZDR) practice, where agents monitor activity per session without saving data. The new system expands this to long-horizon monitoring, analyzing inputs and outputs across multiple conversations to catch bad actors who spread malicious requests over time. If triggered, it sends a narrow signal to OpenAI, which then decides if enforcement is needed, potentially contacting the customer for more context—though sharing data remains optional.

Anthropic, which also generally follows ZDR except for covered models, says any human review of customer data goes through a controlled path with a small set of approved reviewers, and all review sessions are logged in a tamper-proof record.

The timing is notable as competition heats up: OpenAI's Q2 growth slowed relative to Anthropic, whose annualized revenue run rate is reportedly $65 billion. Anthropic investors have floated a possible $2 trillion IPO, while OpenAI is also working on its own public offering.

Source: TechCrunch