OpenAI curbs weapons details for teens, plans safety reports

OpenAI says ChatGPT will give teens less practical information about weapons and security systems, even when a question shows no sign of violent intent. The company announced the change Friday in a post that also commits it to a public safety report twice a year.

OpenAI says it made more than 200 referrals to law enforcement in the third quarter of 2026 over imminent threats of violence.

Limits for every user

Some of the new rules are not just for teens. OpenAI says ChatGPT will give less detail about past mass attacks and the people who carried them out, for all users. That covers reconstructions of attacks, the weapons and equipment used, how they were obtained and comparisons of how much harm was done. OpenAI says the limits apply regardless of a user’s age or reason for asking. People can still ask about how victims and communities were affected and about warning signs, the company says.

When recent chats show warning signs of possible violence, ChatGPT will answer more cautiously, OpenAI says. It says it is working to apply those stricter limits when it predicts a user is under 18, even with no such signs. ChatGPT will also be able to end a chat on its own when someone appears fixated on violence. If ChatGPT ends several chats that way, OpenAI says it may temporarily pause the account.

Teens, age checks and parents

OpenAI estimates a user’s age from signals such as behavior patterns. It says that system now reassesses users continually and moves likely teens into the under-18 version. Adults placed there by mistake can confirm their age through Persona, which OpenAI calls its age verification partner. The post does not say what information Persona collects.

Parents who link accounts can already get alerts when trained reviewers spot a serious self-harm concern or signs of disordered eating, and, since July, when a teen’s account is deactivated for violating OpenAI’s violence rules. OpenAI says it is working on alerts when a teen’s account is paused.

The report

OpenAI says its first Community Safety Report will come in early 2027, covering the second half of 2026, with updates every six months. The post also calls it the Community Safeguards Report. OpenAI says it will include anonymized examples, policy changes and supporting data. The new safeguards begin rolling out in the next few weeks, it says.

Our read: Two of these steps can be checked from the outside. Anyone can ask ChatGPT about a past attack and see whether the details are gone, and parents can see whether alerts arrive. The rest rests on OpenAI’s word. A count of “over 200” referrals says nothing about how many were warranted, how many led to action or how many involved people who had done nothing wrong. A report OpenAI writes, on a schedule OpenAI sets, is not an audit. Age prediction cuts both ways: it guesses age from behavior, and the fix for a wrong guess is proving your age to a third party, Persona, by a method OpenAI calls privacy-protective but does not describe. Auto-ended chats and paused accounts need a clear appeal. The parent alerts are a fair step toward putting families, not a model, in charge. Whether the system actually works will show in the first report’s numbers, if they are specific enough to test.

Also today

OpenAI, Google and Microsoft back a single robots.txt line to opt out of AI training: our story. Britain’s competition regulator opens an investigation into Meta’s terms for publisher content: our story. Google says a test satellite ran an AI model on a data-center chip in orbit: our story. Earlier coverage of ChatGPT for teens: our story.

What’s next

OpenAI says the safeguards roll out over the next few weeks. Its first Community Safety Report is due in early 2027.

Sources: OpenAI