OpenAI announces violence safeguards and tighter teen weapons limits

OpenAI said it will begin rolling out new ChatGPT safeguards for violence-related conversations and tighter limits on weapons information for teens within the next few weeks. The company separately committed to publish its first Community Safety Report in early 2027, covering the second half of 2026.

The product and policy changes are the near-term part of the safety announcement. The report is a later transparency commitment: OpenAI said it will update the report every six months after the first edition.

For all users, ChatGPT will give less detail when answering questions about past mass attacks and their perpetrators. OpenAI said the revised boundaries will restrict reconstructions of attacks, including information about weapons and equipment, how they were acquired, and comparisons that could amplify an attacker’s notoriety. Those limits will apply regardless of a user’s age or stated reason for asking, while still allowing factual discussion of victims, affected communities, warning signs and prevention.

OpenAI also said ChatGPT will respond more conservatively to requests for information that could be misused when recent conversations contain possible warning signs of violence. That builds on short, factual safety summaries introduced in May to carry relevant risk context across separate chats for a limited time.

Additional automated interventions remain under development. OpenAI said ChatGPT will be able to end a conversation when it detects fixation on violence or seriously harming another person, and the company may temporarily pause an account after the system ends multiple chats for that reason. It did not provide the thresholds, error rates or appeal process for those interventions.

The company said it already bans accounts that it determines violated its violent-activity policies. In an earlier description of its enforcement process, OpenAI said automated systems flag potentially concerning activity, trained personnel review it in context, and access may be revoked when reviewers determine that a bannable offense occurred. Those enforcement mechanics and their effectiveness are company claims.

For users estimated to be under 18, OpenAI said it will extend existing limits beyond assistance finding or obtaining weapons. ChatGPT will also restrict practical information about weapons and security systems that could facilitate violence even when a request does not show violent intent. Teens will still be able to ask for help preventing or de-escalating violence, seeking support or connecting with a trusted adult.

The restrictions rely in part on OpenAI’s age-prediction system, which the company says evaluates behavioral and account-level signals and applies the under-18 experience when it estimates a user is a minor. Adults placed there incorrectly can use Persona to verify their age and remove the additional teen protections.

OpenAI also said linked parental-control accounts can receive alerts for specified safety concerns and, since July, when a teen’s account is deactivated for violence-policy violations. The company is developing another notification for cases in which repeated violence discussions lead to a temporary account pause.

OpenAI said it made more than 200 law-enforcement referrals in the third quarter of 2026 and asserted that those referrals were helping disrupt violent plots. The announcement did not provide independent validation of either the count or the effectiveness claim. Separately, OpenAI’s May post reported that its cross-conversation safety context improved safe responses by 52% in harm-to-others tests on GPT-5.5 Instant; that figure came from the company’s internal evaluations, not an independent assessment.

The planned Community Safety Report will describe how OpenAI says its systems detect, investigate and respond to safety issues, with anonymized examples, policy changes, mitigations and relevant data. OpenAI has not yet disclosed the report’s methodology or metric definitions. Its first edition is due in early 2027, with subsequent reports planned every six months.