OpenAI will publish a twice-yearly safety report and locks down weapons content for minors
On October 30, 2026, OpenAI announced a semiannual Community Safety Report on violence prevention and tightened ChatGPT’s answers about weapons for users under 18. The company disclosed more than 200 referrals to law enforcement in Q3 2026, a figure it will now publish with every edition.
October 30, 2026. OpenAI announces a Community Safety Report on violence prevention, to be refreshed every six months. October 30, 2026. The company tightens what ChatGPT says about weapons to users under 18. Q3 2026. It made more than 200 referrals to law enforcement over imminent threats of violence. Why it matters: for the first time, OpenAI is committing to publish moderation numbers on a fixed schedule instead of telling the story anecdotally.
The report: making moderation auditable
The heart of the announcement is not a feature; it is a publishing commitment. OpenAI promises a Community Safety Report focused on violence prevention, explaining how its models detect, investigate, and respond to risk, with anonymized examples, new mitigations, and supporting data.
The first report is due in early 2027 and will cover the second half of 2026. After that, the cadence is fixed: one edition every six months. The move echoes the system cards and model cards the industry already publishes, but applied to its most sensitive front — violence and minors.
The point is auditability. Until now, OpenAI’s moderation described itself in principles and qualitative commitments. A dated, quantified report gives regulators, parents, and researchers a concrete benchmark — and hands OpenAI a measurable obligation to improve from one edition to the next.
More than 200 referrals in three months
The number, released ahead of the report, is the most telling: more than 200 referrals to law enforcement in Q3 2026, for imminent threats of violence spotted on the platform. It is the first time this counter has been made public, and it will become a recurring metric.
OpenAI stresses the “distinct” role AI companies play, alongside law enforcement, families, mental-health professionals, and policymakers. The company says it relies on mental-health and behavior experts to assess hard cases.
Reporting an imminent threat is not new — publishing the volume is. It is exactly what regulators have demanded of platforms accused of operating as black boxes.
New limits on weapons and mass-attack content
Two new restrictions stack on the existing guardrails, and they target different audiences.
For users under 18, ChatGPT will now restrict practical information about weapons and security systems, even when the question looks innocent. The policy already barred helping a minor plan violence or obtain weapons; the extension covers content that could enable violence even with no stated intent.
For every user, regardless of age or reason, ChatGPT will offer less detail about past mass attacks and their perpetrators: reconstructions, the weapons and equipment used, how they were obtained, and comparisons of the scale of harm. The rationale is explicit — such detail can amplify an attacker’s notoriety. Users can still ask about the impact on victims and communities, or about prevention.
The line is fine. OpenAI preserves the ability to discuss these events for understanding — but restricts what can be weaponized. That is an editorial stance, not just a technical filter.
Automatic intervention: ending chats and pausing accounts
The most intrusive part concerns how the model behaves mid-conversation. When recent chats show signs of fixation on violence or intent to seriously harm someone, ChatGPT will be able to end the conversation on its own. If several conversations are cut off for that reason, OpenAI may temporarily pause the account.
The mechanism leans on a strengthened age-prediction system: it reads behavioral signals to estimate whether a user is over or under 18, reassesses continuously, and moves an account into the under-18 experience when the signals fit. An adult who is misclassified can verify through Persona, OpenAI’s age-verification partner, to remove the restrictions.
On the parental-control side, existing alerts — for serious self-harm, and since July 2026 for a teen account deactivated over violence-policy violations — will be joined by a notification when a teen’s account is paused for repeatedly discussing violence.
Age prediction is the quiet linchpin of the entire policy. The restrictions matter only if the model correctly guesses who is a minor, and misclassification cuts both ways: an adult pushed into the under-18 experience loses access to legitimate answers, while a teenager who slips through keeps the looser rules. OpenAI’s answer — continuous re-evaluation from behavioral signals, with Persona as the escape hatch for misclassified adults — is a workable compromise, but its real-world accuracy is precisely what the first Community Safety Report will have to document if the numbers are to mean anything. The same logic applies to the automatic chat endings: a guardrail that fires too eagerly will frustrate legitimate users, while one that fires too late fails its purpose. Calibrating between those two failures is the actual product being built here.
A shift bigger than ChatGPT
OpenAI’s move fits a broader industry trend. Google published a detailed safety report for Gemini 4 Argon; Anthropic documents its safety evaluations before each model deployment. The novelty is no longer publishing a report — it is making one a recurring obligation, on a fixed schedule.
The semiannual cadence is the decisive point. A one-off report can be a communications exercise; a report renewed every six months becomes a statistical series, comparable from one edition to the next. Researchers will be able to measure whether referrals rise, refusals sharpen, and false positives fall.
OpenAI’s Model Spec rewrite plays an infrastructural role here: the five values now structuring it provide the reference frame against which the report’s numbers can be judged. Without that normative base, a raw counter — “200 referrals” — would remain hard to interpret.
The open question is independence. The numbers will be published by the company itself, with no external audit at this stage. That is progress toward transparency, not a guarantee: the report’s credibility will be judged by whether it survives verification from third-party researchers.
The regulatory pressure behind the timeline
The announcement does not land in a vacuum. Earlier in the month, eight US states sued OpenAI, claiming it failed to protect young users. The state of New York passed a law governing AI chatbots aimed at children.
The Community Safety Report slots into that context: publishing numbers on a fixed schedule is a direct answer to the argument that the guardrails are unverifiable. It also extends OpenAI’s recent Model Spec rewrite, which replaced a style checklist with five explicit values.
Read the announcement for what it is: a change in the regime of evidence. OpenAI is moving from quietly enforcing its rules to regularly publishing metrics. Outsiders will soon be able to judge whether the guardrails actually work — rather than taking the promise on faith.
Verdict. If you manage a fleet of users or a product exposed to minors, wait for early 2027 to read the first Community Safety Report: it is the first factual basis for evaluating OpenAI’s moderation. If you are a parent, link your account to your teenager’s to receive alerts. And if you are an adult caught in the under-18 experience, Persona verification removes the limits in a few steps.