Thought Leadership
OpenAI is strengthening data protection for enterprise customers with its new Private Safety Processing feature, which identifies security risks without retaining customer data.
OpenAI has unveiled a preview of a new security feature called Private Safety Processing. The feature is designed for business customers with strict Zero Data Retention (ZDR) requirements. The system analyzes automated patterns across multiple interactions to detect malicious activity early, including cyberattacks that may be planned incrementally over a series of prompts.
The key difference from existing approaches is that neither OpenAI employees nor external reviewers can access the actual inputs or outputs. Data either remains within the customer’s infrastructure or is encrypted on OpenAI’s servers using customer-controlled keys. If suspicious activity is detected, OpenAI receives only an anonymized signal identifying the relevant threat category.
Competitive Edge Over Anthropic’s Data Retention Policy
With this move, OpenAI is seeking to differentiate itself from competitors such as Anthropic in the race for enterprise customers. Anthropic recently announced that it would retain interaction data for up to 30 days for security reviews involving models in its Mythos class, including Fable 5.
For many organizations in highly regulated industries such as financial services and healthcare, such a retention period can effectively rule out the use of these systems. OpenAI, by contrast, aims to enable security reviews while maintaining strict ZDR agreements.
Pressure From Security Risks and Regulation
The privacy push comes amid growing concerns about incidents involving autonomous AI agents and models escaping their intended testing environments. OpenAI recently paused training on new frontier models after their capabilities advanced faster than the safeguards designed to control them.
Beyond the technical risks, the company is also responding to potential legal consequences and increasing pressure from enterprise customers demanding maximum control over their data.
(Editorial Team)
