OpenAI unveils Private Safety Processing to challenge Anthropic on enterprise privacy

OpenAI has previewed a new service called Private Safety Processing, an automated system designed to monitor for potential artificial intelligence misuse across multiple sessions without retaining customer data. The announcement represents a direct competitive response to Anthropic, which recently introduced a policy allowing it to retain user data for 30 days for specific "covered models" to facilitate safety reviews.

The new technology extends Zero Data Retention (ZDR) capabilities by analysing inputs and outputs across conversations to detect malicious activity. OpenAI states that the system uses an agent to monitor for abuse without human review of user content. If the system is triggered, it sends a "narrowly defined signal" to OpenAI, which may then contact the customer for context or to determine if enforcement is necessary.

This approach allows OpenAI to detect malicious use, such as malware engineering spread across multiple sessions, without retaining or inspecting customer data. A spokesperson told TechCrunch that the new tech helps detect malicious use of AI that takes place over multiple sessions, where a bad actor might spread out their requests to avoid detection.

Anthropic’s contrasting policy, announced in July, retains user data for 30 days for "covered models," including Mythos-class and Fable models, to allow for safety analysis. The move has concerned some enterprise clients that handle large amounts of sensitive data. Anthropic notes that human review of customer data can occur, but only through a "controlled access path" involving a small set of approved reviewers, with all sessions recorded in a tamper-proof log.

The announcement highlights intensifying competition between the two AI firms, with both companies seeking advantages in the enterprise market. Recent reports indicate that OpenAI’s Q2 growth was slower than Anthropic’s, with Anthropic’s annualised revenue run rate reportedly at $65 billion.

Both companies are reportedly working towards initial public offerings, with Anthropic investors suggesting a potential valuation of $2 trillion. The rivalry underscores the delicate balance AI companies must strike between respecting enterprise customer privacy and monitoring usage for potential issues.