OpenAI announced the launch of the "Private Safety Processing" service, available for preview to select customers. This service monitors potential misuse of AI models through automated systems without retaining customer data, further expanding its zero data retention (ZDR) security capabilities.
As AI model capabilities rapidly improve, the likelihood of models being used in risk scenarios such as cyberattacks and malicious code generation increases. AI companies are facing challenges in balancing customer privacy with security regulations. OpenAI's newly launched Private Safety Processing aims to address the issue where enterprise customers are unwilling to share sensitive data but still require security protection.
According to the introduction, this technology is an upgraded solution based on OpenAI's existing ZDR policy. Traditional ZDR detects abnormal usage behavior through API proxy programs per session, avoiding the storage of customer data. Private Safety Processing further supports analyzing inputs and outputs across multiple conversations, with agents monitoring potential risks over time. When abnormal activity is detected, related interaction information is captured and analyzed for correlation to identify distributed malicious requests.
OpenAI stated that the system can detect patterns of malicious use across sessions without human review of user conversations. For example, attackers may split requests to bypass single-session detection, while this technology can determine whether there are risks such as creating malicious software or launching cyberattacks through multi-turn interactions.
When the system triggers a security alert, OpenAI may receive a clear risk signal and decide whether to take further action based on the situation. If intervention is needed, OpenAI will contact the customer to understand the context, and the customer can independently decide whether to provide more data to OpenAI.
This solution is also seen as a response from OpenAI to Anthropic's recent data retention policy. Anthropic previously announced a policy allowing it to retain user data and conversation records for 30 days when processing certain "protected models," raising concerns among some enterprise customers about the security of sensitive information. However, Anthropic stated that customer data reviews are conducted only through controlled access paths, performed by a small number of approved reviewers, and complete tamper-proof logs are recorded.
Currently, OpenAI and Anthropic are engaged in fierce competition regarding model capabilities, security systems, and commercialization. Market data shows that Anthropic has experienced rapid growth, with reported annual revenue reaching $65 billion and being favored by investors to proceed with an IPO valued at approximately $2 trillion; at the same time, OpenAI is also preparing for an IPO. As enterprise-level AI applications expand, privacy protection and security governance capabilities have become important competitive dimensions for AI companies.