If your AI agents call a public LLM, “delete” isn’t fully in your control: lessons from the 2025 OpenAI preservation order

In May 2025, during the New York Times copyright case, a federal magistrate judge ordered OpenAI to preserve output log data that would otherwise be deleted. That included chats users had deleted and Temporary Chats. OpenAI described the order as also covering API content, and said ChatGPT Enterprise was excluded. Users weren’t parties to the case, so they couldn’t object. The blanket order ended on September 26, 2025, but logs already preserved stayed preserved (apart from data from users in the EEA, Switzerland, and the UK). In January 2026, a judge affirmed an order requiring OpenAI to produce a sample of 20 million de-identified logs. OpenAI’s stated policy is that deleted data is removed within 30 days unless it is legally required to retain it. For part of 2025, that exception applied broadly. Why this matters for agent builders: agents usually touch more data than a chat window does, such as CRM records, documents, tickets, and customer messages, and they send it to a public LLM on every call. If that provider ends up under a legal hold, how long everything your agent sent is retained is out of your hands, even if the case has nothing to do with you. That’s one reason some teams run agents on private models in their own environment (VPC or on-premise), where they set retention, access, and audit rules. It isn’t automatic safety, since you still have to manage logs, memory stores, and backups. And for low-stakes workflows, a public LLM is often faster to ship. Questions for people building agents: Do you treat provider retention terms as a design constraint when you pick a model? Do you log full prompts and tool outputs in your agent’s own traces? If so, what’s your purge policy? For clients with sensitive data, do you use public LLM APIs with zero-retention terms, or do you move to private models? Disclosure: I work in enterprise AI agents and private deployment, so I have a bias on this topic.