OpenAI unveiled a new automated monitoring system that detects potential abuse of its AI models while storing none of the customer's information, according to an announcement this week. The company is rolling out Private Safety Processing to select customers as a preview, positioning the technology as a direct counter to rival Anthropic's data-retention approach. The system represents OpenAI's latest attempt to balance enterprise privacy demands with safety oversight as AI capabilities expand.
The new technology extends OpenAI's existing Zero Data Retention policy by analyzing inputs and outputs across multiple conversations rather than single sessions. An automated agent monitors these interactions and, when triggered, examines them for signs of potential misuse without human reviewers accessing customer conversations. If the system detects concerning activity—such as a bad actor spreading malware engineering requests across sessions to evade detection—it transmits what the company describes as a "narrowly defined signal" to OpenAI warning of specific activity types. The company can then determine whether enforcement actions are warranted and may contact the customer for additional context, with customers retaining discretion over whether to share data with OpenAI.
This approach contrasts sharply with Anthropic's policy announced in July, which permits the AI lab to retain user data—including all sessions and their conversations—for 30 days for "covered models" such as Mythos-class models and future models with comparable capabilities. The policy has frustrated some enterprise customers handling sensitive information who object to having their data stored or examined by the AI lab. While Anthropic also largely follows Zero Data Retention principles, it makes an exception for covered models like Fable. When human review of customer data occurs at Anthropic, it happens only through a controlled access path involving a small set of approved reviewers, with every review session recorded in a tamper-proof log that reviewers cannot alter or suppress.
The competing privacy strategies emerge as both companies jostle for enterprise market share in an increasingly tense rivalry. A recent report showed that OpenAI's second-quarter growth lagged behind Anthropic's, with Anthropic's annualized revenue run rate now reportedly reaching $65 billion. AI companies face mounting pressure to install safety guardrails against model misuse as capabilities grow more powerful, yet they must simultaneously respect corporate clients' privacy requirements. OpenAI's Private Safety Processing attempts to thread this needle by enabling what the company characterizes as long-horizon safety monitoring that catches malicious patterns spanning multiple sessions—behavior a single-session scan would miss—without retaining the underlying customer data. Investors have speculated Anthropic could pursue an initial public offering at a $2 trillion valuation, while OpenAI is also working toward going public. The privacy-versus-safety debate will likely shape how enterprises choose between AI providers, particularly for organizations managing highly confidential information that can't tolerate third-party data retention. Companies prioritizing zero-knowledge architectures may gravitate toward OpenAI's approach, while those willing to accept temporary data storage for enhanced safety review might find Anthropic's model acceptable.