A chat that ends itself
OpenAI said on 30 October that ChatGPT will be able to end a conversation on its own when it shows signs that someone is fixating on violence or on seriously harming another person. If the system ends several chats for that reason, OpenAI says it may temporarily pause the account.
That is a different kind of intervention from a refusal. A refusal declines one request and leaves the conversation running; ending the chat removes the thread the user was building in. OpenAI frames it as interrupting a pattern “in the moment”, and says it is working with mental health and violence prevention experts on when and how to intervene.
The number in the post
The announcement also discloses a figure OpenAI has not published in this form before: “we made over 200 law enforcement referrals in Q3 2026.” Those are cases where the company judged a threat of violence to be imminent and took it to police.
OpenAI is committing to a Community Safety Report that will cover violence prevention and how well its safeguards work, with anonymised examples, new mitigations and supporting data. The first edition is due in early 2027 and will cover the second half of 2026, with updates every six months after that. So the 200 referrals are a preview of a reporting cycle that has not started yet.
Tighter limits on describing past attacks
The policy change that applies to everyone concerns past mass attacks. OpenAI says it is restricting how much detail ChatGPT gives about those events and the people who carried them out — specifically reconstructions that include the weapons and equipment used and how they were obtained, and detailed comparisons of scale that could amplify an attacker’s notoriety.
The company is explicit that these limits apply “regardless of a user’s age or reason for asking”, and explicit that the material is often available elsewhere online. It also says people should still be able to ask about how victims and communities were affected, and about recognising warning signs.
Separately, when recent chats show warning signs of possible violence, ChatGPT is to respond more conservatively to later requests that could be misused — the company described the underlying mechanism, short factual summaries of earlier warning signs carried into a later chat, in May.
Teens get the stricter default
For users under 18, OpenAI is restricting practical information about weapons and security systems even when a question appears harmless and shows no sign of violent intent. The same stricter handling is to be applied when the company predicts a user is under 18, whether or not they have said so.
That prediction is doing a lot of work. OpenAI says it has strengthened its age prediction system to assess new signals continuously as people use ChatGPT, and to move teens into the under-18 experience when appropriate. Adults placed there incorrectly can verify their age through Persona, the company’s verification partner.
What to watch
The first Community Safety Report, in early 2027. A referral count is a measure of activity, not of accuracy, and the questions it raises are the ones the report would have to answer: how many referrals led to anything, how often the system was wrong, and how often an account pause hit someone who was not a threat. OpenAI has promised anonymised examples and relevant data. Which data is the part still to be decided.
