OpenAI plans to restrict practical guidance about weapons and security systems for teens using ChatGPT, including when their questions show no apparent intent to commit violence. The company says the safeguards will start reaching users over the coming weeks, with additional limits on descriptions of mass attacks applying to users of every age.
For teens, the changes extend existing rules against assistance with finding or acquiring weapons to information that could help someone commit violence. OpenAI says teens will still be able to ask for help calming violent situations, finding support and reaching an adult they trust. The company is also developing more restrictive responses for users it estimates are younger than 18, without requiring warning signs in their conversations.
Read more: OpenAI Expands ChatGPT Ads to 26 More Countries, Reaching 89
Across all age groups, OpenAI plans to reduce the detail ChatGPT provides about previous mass attacks and their perpetrators. Restrictions will cover accounts of how attacks unfolded, the equipment involved and its acquisition, plus detailed comparisons of harm that could increase an attacker’s notoriety. OpenAI says users should still be able to explore the effects on victims and communities and ask about prevention.
OpenAI is also developing automatic interruptions for conversations that suggest a user is becoming preoccupied with violence or seriously injuring someone. Repeated interruptions could lead to a temporary account pause, and the company is working on alerts to parents when that happens to their teen. Linked parental accounts already support notifications when a teen’s account is deactivated over violence policy violations, a capability added in July.
The company says its age prediction system now keeps evaluating fresh signals during use and can move users into its teen experience. Adults mistakenly assigned those protections can verify their age through Persona to remove them. OpenAI also plans to use warning signs from recent conversations to make subsequent responses more cautious about potentially misused information.
OpenAI says it referred more than 200 cases to law enforcement in Q3 2026. It plans to publish its first Community Safety Report in early 2027, examining July through December 2026, then issue updates at six-month intervals. The planned reports will include anonymized cases, safeguard changes and data about its response to potential harm.
The user-facing restrictions extend OpenAI’s safety work beyond the frontier reinforcement learning training addressed in guidelines published September 28. Those earlier guidelines excluded deployment alignment properties, and their proposed safety case remained an aspiration rather than a requirement.
