OpenAI previews cross-session safety systems for autonomous AI agents

OpenAI has announced it is previewing new safety measures aimed at long-running autonomous AI systems, a growing category of agents that operate over extended periods with minimal human supervision. Unlike single-turn models, these agents can chain together many actions, such as managing emails, updating databases, or orchestrating business workflows. The new safety systems are designed to detect risks that only become apparent across related interactions, rather than evaluating each action in isolation. This includes scenarios where an agent might gradually drift toward harmful behavior or where a series of otherwise benign steps could cumulatively cause data leaks or operational damage.

OpenAI also confirmed it will continue offering Zero Data Retention (ZDR) for its frontier models, a key feature for enterprises that need strict privacy guarantees. By combining ZDR with cross-interaction safety monitoring, OpenAI aims to make autonomous agents more viable for regulated industries like finance, healthcare, and legal services. The preview signals a shift from single-prompt AI safety toward continuous, context-aware oversight. While no specific product name was given, the announcement positions OpenAI as addressing one of the biggest blockers to deploying AI agents in production: trust. The company is likely to integrate these safety systems into its API and enterprise offerings, making long-running agents more predictable and auditable for businesses that want to scale automation without sacrificing control.