OpenAI Unveils 'Zero Data Retention,' Targeting Anthropic's 30-Day Policy

OpenAI previewed a new service, Private Safety Processing, allowing eligible enterprise and API customers to maintain zero data retention while using frontier models. The system can check for potential abuse across multi-turn conversations, but OpenAI employees cannot see customer prompts or model responses. Customer data can remain on their own servers, or be stored with OpenAI but encrypted using the customer's own keys, which OpenAI cannot access.

If the system detects a risk, it only returns a limited safety signal to OpenAI, such as the type of suspicious activity detected, without handing over the original conversation. This directly targets a controversial point for Anthropic. Fable 5 and Mythos 5 require retaining user inputs and model outputs for 30 days to check for attacks across multiple requests.

Microsoft previously restricted employee use of Fable 5 due to this policy. Anthropic argues that retaining data is a necessary cost for effective safety monitoring, while OpenAI now aims to prove both can be achieved simultaneously. Private Safety Processing is currently being tested with a small number of customers, with OpenAI planning to launch it in September and publish a technical whitepaper. Regular ChatGPT Free, Plus, Go, and Pro users are not included in this ZDR scheme, and existing data settings will not change as a result.

[BlockBeats]