OpenAI Faces Access Revocations in Cybersecurity Program, Unveils New Privacy-Focused Safety System

Multiple security researchers reported losing access to OpenAI’s Trusted Access for Cyber program, which grants vetted researchers reduced-restriction access to advanced AI models for legitimate security work. OpenAI confirmed the issue stemmed from a technical error affecting a limited number of users, primarily those located outside the U.S. and Europe, and asked affected researchers to reapply for verification. The company said access to Daybreak Blue, its latest researcher tier launched August 10, was disrupted for select accounts, though the underlying cause remains unclear.

Separately, OpenAI introduced Private Safety Processing, a new privacy-focused monitoring system designed to detect misuse across multiple conversations without retaining customer data. The system builds on OpenAI’s existing Zero Data Retention approach, using automated agents to flag suspicious activity spanning multiple sessions, then notifying OpenAI only through narrowly defined signals if intervention becomes necessary.

The announcement comes amid growing rivalry between OpenAI and Anthropic, particularly after Anthropic introduced a data retention policy allowing limited human review of certain “covered models” under controlled, logged conditions. OpenAI’s approach positions itself as a more privacy-preserving alternative. The competitive dynamic unfolds as both companies pursue eventual public offerings, with Anthropic recently reporting a $65 billion annualized revenue run rate and OpenAI continuing preparations toward its own IPO.