OpenAI’s safety communications have so far been mostly defensive: here is the classifier, here is the red-team report, here is how we detect misalignment. The new announcement is different in kind, because it commits the company to publishing what its safeguards fail to catch.
The Report
OpenAI will launch a Community Safeguards Report on violence prevention and other community safety issues, covering how its systems are trained to detect, investigate and respond, what violence-related harm looks like in practice, and which mitigations follow. The first edition covers the second half of 2026 and lands in early 2027, with updates every six months thereafter.
The context is the number OpenAI chose to lead with: over 200 law enforcement referrals in Q3 2026, which the company links to disrupted violent plots.
Duller Answers, For Everyone
Alongside the report, OpenAI is tightening how ChatGPT responds to questions about past mass attacks and their perpetrators — restricting reconstructions that detail weapons, equipment, how they were obtained, and comparisons of scale that could amplify an attacker’s notoriety.
Two details matter. The limits apply regardless of the user’s age or stated reason for asking. And the framing is explicitly not a ban on the topic: users should still be able to understand these events, learn about warning signs and prevent violence — just not rebuild them.
Warning Signs and Teens
The context-sensitivity work from May is being extended. When recent chats show warning signs of violence, ChatGPT will respond more conservatively about information that could be misused — and the stricter limits will also apply whenever the system predicts a user is under 18, even with no violent intent at all. Where a conversation shows fixation on harming someone, ChatGPT can now end the chat automatically; repeated endings can mean a temporary account pause.
For teens, the limits extend to practical information about weapons and security systems that could facilitate violence, again regardless of intent. Age prediction, launched last year, is being strengthened to re-evaluate continuously and move users into the under-18 experience as signals accumulate.
Related: OpenAI announcement