OpenAI's Private Safety Processing monitors across sessions without retaining customer data

OpenAI is previewing a new safety service, Private Safety Processing, for select customers. The automated system watches for AI misuse while retaining none of the customer's data, building on the Zero Data Retention (ZDR) policy OpenAI already uses within its API. ZDR typically monitors for abuse on a per-session basis, but Private Safety Processing expands this to long-horizon safety monitoring: an agent assesses inputs and outputs across multiple conversations to identify patterns of malicious behavior. This helps catch bad actors who deliberately spread their requests across many sessions, such as someone trying to engineer malware for a cyberattack. If the system is triggered, it sends only a narrow signal to OpenAI about the specific type of activity detected. OpenAI then decides whether enforcement is necessary and may reach out to the customer, who can choose whether to share additional data at their discretion.

This move directly contrasts with Anthropic's recently announced data-retention policy, which stores user sessions and conversations for 30 days for covered models — including all Mythos-class models and future models with similar capabilities, like Fable. Anthropic says any human review would happen only through a controlled access path involving a small set of approved reviewers, with every session recorded in a tamper-proof log. Still, enterprise customers handling sensitive data have been concerned about having their interactions harbored or inspected. The tension comes amid intense competition: recent reports show OpenAI's Q2 growth trailed Anthropic, whose annualized revenue run rate has reportedly reached $65 billion. Anthropic investors have floated a potential $2 trillion IPO, while OpenAI is also working on its own IPO.