OpenAI announced two things on Thursday: a recurring report on violence prevention, and narrower limits on weapons content for minors. The Community Safety Report is due in early 2027, will cover the second half of 2026, and will then be updated every six months.
Along the way comes a number that had not been published anywhere: OpenAI made more than 200 law enforcement referrals in Q3 2026.
What changes for every user
ChatGPT will give less detail about past mass attacks and the people who carried them out, even where the same information is available elsewhere online. Specifically restricted: reconstructions of past attacks, details about the weapons and equipment used and how they were obtained, and comparisons of scale of harm that could amplify an attacker’s notoriety. Questions about context, impact on victims and communities, and recognizing warning signs stay open. The limits apply regardless of a user’s age or stated reason.
The chat that stops on its own
The most interesting part technically sits in the section on warning signs. When a conversation shows that someone is fixating on violence or on seriously harming another person, ChatGPT will be able to end the chat itself, to interrupt the pattern in the moment. If it ends several chats for that reason, OpenAI may temporarily pause the account. And where recent chats contained warning signs, the model answers more conservatively later, using the short factual summaries OpenAI introduced in May.
The same stricter limits are meant to kick in when the system predicts a user is under 18, even with no sign of violent intent at all.
Teen safety moves upstream
The old rule was straightforward: no help planning violence, no help obtaining weapons. The new part is the step before that. Practical information about weapons and about security systems is now off limits for teens even when the question looks harmless and nothing suggests intent. Prevention, de-escalation and asking for help all remain available.
Age prediction, which OpenAI launched last year, now keeps reading behavioral signals as people use ChatGPT and moves teens into the under-18 experience when it fits. Anyone placed there by mistake can verify as an adult through Persona, OpenAI’s verification partner. Parental controls gain another alert: parents will be notified if a teen’s account is paused for repeatedly discussing violence.
A model that closes conversations is new behavior
Content limits are familiar. Account bans are familiar. A model that shuts down a live conversation by itself is something else: it acts on the session, not on the answer. If you read ChatGPT’s behavior as a reference point for your own products, that is the bit worth noting — and the open question is false positives. The post only says OpenAI is still working on recognizing warning signs reliably.