OpenAI is preparing to launch a system called private safety processing in September, designed to monitor its AI models for dangerous behavior while keeping user data shielded from human eyes, including its own employees. The company has already implemented security measures where human data access is limited to specific severe cases, along with 30-day data retention limits on the enterprise side. OpenAI recently paused certain training activities for frontier models due to potential risks, signaling that its preparedness framework is being applied with real consequences. No public confirmation exists regarding the September launch and implementation details, including what gets flagged, who reviews content and how long data is retained, remain to be determined.
Source: Read the original article
