OpenAI previews cross-interaction safety tech, pledges not to keep customer data

OpenAI announced it is previewing, for select customers, technology that detects patterns and safety risks across multiple model interactions rather than monitoring each interaction in isolation. The explicit goal is to ensure proper tool use and catch emergent risky behavior in agentic deployments—while, critically, not retaining customer data. OpenAI framed the no-retention promise as a direct competitive pitch against Anthropic on enterprise privacy grounds.

The cross-interaction approach is the same architectural shift underlying OpenAI's post-Hugging-Face security overhaul: monitoring that watches sequences of agent actions for attack patterns invisible in any single call. Packaging it as a privacy differentiator—'we watch behavior, not your data'—is the clever part, letting OpenAI claim both stronger safety and stronger privacy simultaneously.

The timing is pointed. Anthropic's Boris Cherny spent the same week emphasizing that Claude Security customers 'own and control their own data and Anthropic retains none,' and Microsoft's reliance on OpenAI as an Azure subprocessor remains an EU data-residency watch item. Enterprise buyers are increasingly making data-retention and residency the deciding procurement criteria—the same logic France used to exclude OpenAI from government security work.

Separately, OpenAI confirmed a login and signup outage affecting ChatGPT users, a reminder of reliability pressures at scale. The skeptical question for the safety tech: can cross-interaction monitoring genuinely detect risk without retaining enough data to reconstruct behavior? Watch whether OpenAI publishes technical details that let enterprises verify the no-retention claim.