OpenAI announced a new enterprise safety system called Private Safety Processing, positioning it as a privacy-centric alternative for monitoring customer misuse. The system uses automated agents to detect potential abuse across customer conversations without storing any customer data. According to TechCrunch, the announcement directly challenges Anthropic's data retention policy, which stores customer conversations and sessions for 30 days on its Mythos-class models and "future models with similar capabilities."
Private Safety Processing extends OpenAI's existing Zero Data Retention policy, which currently monitors individual sessions for abuse without persisting data. The new system addresses a critical limitation by analyzing patterns across multiple conversations over extended periods, enabling detection of coordinated misuse campaigns. An example is detecting an actor spreading requests across separate interactions to incrementally engineer malware or stage cyberattacks. When the automated system detects suspected abuse, it sends OpenAI a narrowly defined signal describing the specific activity type, allowing the company to determine whether to contact the customer or take enforcement action.
OpenAI's zero-retention stance directly contradicts Anthropic's approach. TechCrunch reported that Anthropic retains customer data for 30 days on covered models to support safety analysis. Anthropic contends the retention is necessary and emphasizes that human review occurs only through controlled access paths involving a small set of approved reviewers, with all access recorded in tamper-proof logs that reviewers cannot suppress or modify. Enterprise customers holding sensitive data have nonetheless raised concerns about Anthropic's retention practice.
The privacy and safety policy competition reflects broader market rivalry between OpenAI and Anthropic, both pursuing enterprise customers ahead of planned public offerings. Anthropic's annualized revenue run rate reached $65 billion, according to TechCrunch, with investors projecting potential IPO valuations near $2 trillion. OpenAI is pursuing similar IPO plans while competing directly for enterprise sales.