OpenAI launches cross-conversation safety monitoring without retaining data

OpenAI introduced Private Safety Processing, an automatic system that monitors misuse across multiple separate conversations while storing no customer data. The move directly answers Anthropic’s retention policy, which since July has allowed the lab to keep whole sessions for thirty days in models classified as “covered”, including all Mythos and future models with similar capabilities.

Anthropic’s policy was intended to enable retroactive scanning for suspicious behavior, but organizations handling sensitive data raised concerns: the data remains with the provider, exposed to human review, without client control over timing or scope. OpenAI already runs a Zero Data Retention (ZDR) policy that monitors each session individually via agents inside the API, without human intervention. Anthropic follows the same principle except for the “covered” models like Fable, where the thirty-day exception is retained.

Private Safety Processing extends ZDR to a longer horizon: the agent examines inputs and outputs from several conversations together, not just one, and looks for patterns indicating abuse that unfolds over time. A classic example is an attacker splitting requests to generate malware across multiple sessions to evade point-in-time detection. When the system flags a suspicion, it sends a “well-defined signal” to OpenAI, which decides whether enforcement is required; only then does the company contact the client for context, and the client may voluntarily share information.

Anthropic emphasizes that human review occurs only through a controlled access path, with a small group of authorized reviewers, and each review session is logged in an immutable audit trail that reviewers cannot delete or alter. Both approaches aim to detect malicious use without compromising privacy, but the architectures differ: cross-session automation versus documented human oversight.

The rivalry for enterprise customers is intensifying. Recent reports show OpenAI’s second quarter grew more slowly than Anthropic, whose annual revenue is estimated at 65 billion dollars. Anthropic’s investors are discussing an IPO at a valuation of two trillion dollars, and OpenAI is pursuing a similar path. Private Safety Processing is another tool in the battle for enterprise trust, which will determine where the large capital flows.