OpenAI made more than 200 referrals to law enforcement in the third quarter of 2026, the company disclosed on October 30, 2026, in a post that also set out new limits on how ChatGPT discusses past mass attacks, tighter weapons restrictions for users it believes are under 18, and a commitment to publish a Community Safety Report every six months. The first report is due in early 2027 and will cover the second half of 2026. According to OpenAI, the safeguards described in the post "will begin rolling out in the next few weeks."

In Short

Between July and September 2026, OpenAI told police or other authorities about possible violence it spotted on ChatGPT more than 200 times. Soon ChatGPT will give everyone less detail about past mass attacks, give teens less practical information about weapons, and be able to end a chat when someone keeps fixating on hurting another person. OpenAI also promised a public safety report every six months, starting in early 2027, so outsiders can see how well these protections work.

A number without a baseline

The referral figure sits in the second paragraph of the post and comes with almost no detail. OpenAI states that "we made over 200 law enforcement referrals in Q3 2026" and that "these referrals are making a positive difference in disrupting violent plots." The post does not say which countries or agencies received the referrals, how many related to threats against specific people as opposed to broader plots, how many led to arrests or charges, or how the third-quarter total compares with any earlier quarter. It does not give the number of ChatGPT users or conversations the referrals were drawn from.

OpenAI supports that claim with links to two news reports. One, published on September 21, 2026, by the Kyrgyz news site open.kg and citing the BBC, describes the arrest of a 36-year-old farmer in the Brazilian state of Espírito Santo who was accused of discussing, over two months, plans to kill his eight-year-old son and attack police officers. According to that report, OpenAI passed information about the user to the FBI, which forwarded it to Brazil's Ministry of Justice. The man was held for 54 days and released on August 13 with an electronic bracelet and a ban on approaching his son; his lawyer denies guilt and argues that OpenAI handed investigators only his text queries, without what ChatGPT generated in reply. The second link points to a report by WPBF, a Florida television station, about a cyberstalking case in South Palm Beach.

OpenAI described its referral policy in an April 28, 2026 post titled "Our commitment to community safety." In that document the company said it notifies law enforcement when conversations indicate an imminent and credible risk of harm to others, that its criteria are flexible enough that a user need not spell out a target, means and timing, and that automated flags are reviewed in context by trained staff. The October post adds that OpenAI continues "to rely on mental health and behavioral experts to help us assess difficult cases."

The company frames its position cautiously. "We recognize AI companies have a distinct role to play alongside law enforcement, family members, mental-health professionals, and policymakers, who bring critical authority, context and real-world reach," the post reads.

A report every six months

The most durable commitment in the post is a reporting schedule. OpenAI says it will launch a Community Safety Report "on violence prevention and other community safety issues, and the effectiveness of our safeguards." The report will explain how OpenAI trains its systems to detect, investigate and respond to such issues, and will include "anonymized examples of how violence-related harm manifests and what we're learning," together with new mitigations, policy changes and "relevant data to support these findings."

The first edition is promised for early 2027, covering July to December 2026, "with updates every six months thereafter." No month is given. The post does not list the metrics the report will contain: there is no undertaking to publish referral counts by country, account bans, the number of chats ended automatically, or error rates for the age prediction system described later in the same post. OpenAI's stated commitment is "to share meaningful findings about how we identify, review, and respond to potentially harmful content," which leaves the company to decide what counts as meaningful.

Even the report's name is unsettled. The opening paragraph announces "a new Community Safeguards Report," while every later reference, including the section heading, calls it the Community Safety Report.

Transparency reports of this kind have already found their way into legislation. Illinois SB 3444, introduced on February 4, 2026, would shield developers of frontier AI systems from civil liability for mass-casualty harms provided they did not act intentionally or recklessly and had published a safety protocol and a transparency report. OpenAI testified in favor of the bill; Anthropic opposed it. The Illinois text uses a compute threshold to define a frontier model, and Wired reported on April 9, 2026, that the threshold would likely cover OpenAI, Google, xAI, Anthropic and Meta.

Past mass attacks: new limits for every user

The first policy change applies regardless of age. "Our models are trained to not give instructions that would help someone plan or carry out violence," the post states. OpenAI is now going further on a narrower category: questions about past mass attacks and the people who carried them out.

ChatGPT will be directed to provide less detailed information in those areas "even if it is available from other online sources." That clause sets aside a familiar defense of permissive answers: that repeating information already published elsewhere adds no harm. According to OpenAI, the tightened policy restricts reconstructions of past mass attacks, "including details about the weapons and equipment used and how they were obtained, and detailed comparisons on scale of harm that could amplify an attacker's notoriety." OpenAI adds: "These limits apply regardless of a user's age or reason for asking."

What remains permitted is described in broad terms. Users should still be able to ask ChatGPT to help them understand such events, how victims and communities were affected, how to recognize warning signs and how to prevent future violence. Where does understanding an attack end and reconstructing it begin? The post does not define the line further, and it gives no examples of prompts on either side of it.

The pressure on this point has been building in public. At a New York City Council hearing on October 5, 2026, Imran Ahmed of the Center for Countering Digital Hate testified that his organization's tests had found eight of ten chatbots would regularly help plan violent attacks. Florida's lawsuit against OpenAI and Sam Altman, filed on June 1, 2026, cites the April 2025 shooting at Florida State University, in which two people were killed and six injured, and a separate federal suit filed on May 10, 2026, which alleges more than 200 exchanges between the accused gunman and ChatGPT. Those claims are allegations that have not been tested in court.

Warning signs that build across conversations

The second change deals with context that accumulates over time. "Sometimes a request that seems ordinary on its own takes on a different meaning in light of an earlier conversation," the post observes.

OpenAI points back to work it published in May. According to the company's May 14, 2026 post, ChatGPT can now draw on safety summaries: short, factual notes about earlier safety-relevant context, produced by a model trained for safety reasoning, narrowly scoped, kept only for a limited time and used only when a serious safety concern arises. That post focused on self-harm and harm-to-others scenarios. It reported improvements in safe responses of 50% for suicide and self-harm and 16% for harm to others in long single conversations, and of 52% for harm to others and 39% for suicide and self-harm across multiple conversations on GPT-5.5 Instant. OpenAI did not say whether those were relative gains or percentage-point gains.

The October post extends that mechanism. When recent chats show warning signs of possible violence, ChatGPT will "respond more conservatively when sharing information that could be misused." OpenAI says it is also working to apply the same stricter limits whenever "we predict someone is under 18, even if their chats show no signs of violent intent." For predicted minors, in other words, the conservative mode would become the default rather than a response to a signal.

Ending chats and pausing accounts

Enforcement is moving from the account level into the conversation itself. OpenAI already bans accounts that it determines have violated its policies on violent activity. It is now working with mental health and violence prevention experts on "additional, automated ways to intervene earlier."

When a conversation shows signs that someone is fixating on violence or on seriously harming another person, ChatGPT "will be able to automatically end the chat to interrupt that pattern in the moment." Repetition carries a further consequence: "If ChatGPT ends multiple chats for that reason, we may temporarily pause the account."

Several operating details are missing. The post does not say how many ended chats trigger a pause, how long a pause lasts, whether the user is told why a chat ended, or whether a paused account can appeal. The word "may" leaves the pause discretionary. OpenAI calls the work "an ongoing effort" and says it continues to invest in how to recognize warning signs reliably "and determine when and how to intervene."

Automatic termination has been a recurring demand in litigation against the company. The complaint filed by the parents of 16-year-old Adam Raine on August 26, 2025, alleges that OpenAI's moderation flagged 377 of their son's messages for self-harm and that no safety protocol activated to end the conversations. In a suit filed on June 11, 2026, Kristie Carrier, whose 24-year-old daughter Alice died in July 2025, asked the court for relief that includes automatic conversation termination. Both cases concern self-harm rather than violence against others, and the allegations in them have not been proven.

Teens: weapons and security systems

The teen-specific change is the one OpenAI chose to headline, although the wording shifts between sections. The introduction promises "more restrictions around firearms content for minors." The section that describes the change never uses the word firearms. It speaks instead of "weapons-related information" and, more broadly, of "weapons and security systems."

According to OpenAI, ChatGPT for Teens applies age-appropriate safeguards to users under 18, with stronger protections and extra limits on sensitive content. The company's policies already prohibit helping anyone plan or carry out violence and restrict helping teens find or obtain weapons. The new rule goes beyond acquisition: OpenAI is extending protections to "restrict practical information about weapons and security systems that could be misused to facilitate violence, even when there's no sign of violent intent." The limits apply "even when a question appears harmless."

The inclusion of security systems is notable and unexplained. The post gives no examples, so it is not clear whether the restriction covers questions about alarm systems, school or building access controls, surveillance cameras, or something else. Teens, OpenAI says, can still use ChatGPT to help prevent and de-escalate violence, to seek support and to connect with a trusted adult.

At the October 5 council hearing in New York, OpenAI's representative Morgan Dwire told council members that ChatGPT for teens serves users aged 13 to 17, is safe by default and comes with optional parental controls. Meta AI is also available from 13, while Anthropic said Claude is limited to adults.

Age prediction, now continuous

Every teen safeguard depends on knowing, or guessing, who is a teen. "We launched age prediction last year," the post states, describing a system that "analyzes multiple signals such as behavior patterns to estimate whether a user may be under or over 18."

The dating is loose. OpenAI's own January 20, 2026 post, "Our approach to age prediction," described the rollout on ChatGPT consumer plans as underway at that point, with the European Union to follow in the coming weeks. The company's September 29, 2025 announcement of parental controls had described the age prediction system as still being built. The January post listed the signals the model uses: how long an account has existed, the times of day a user is typically active, usage patterns over time and the user's stated age. It also said that when OpenAI is not confident about a user's age, the system defaults to the safer experience.

The October change is about frequency. According to OpenAI, the system has been strengthened to "continually assess new signals as people use ChatGPT and move teens into the under 18 experience when appropriate." An account classified as adult at one point can therefore be reclassified later, based on behavior.

Misclassification is handled through a third party. Adults wrongly placed in the under-18 experience "will always have a fast, privacy-protective way to confirm they are adults and have additional teen protections removed through Persona, our age verification partner." The January post said that confirmation is done with a selfie. The October post gives no accuracy rate, no false-positive rate, no count of accounts moved into the teen experience, and no figure for how many adults have used Persona to reverse a classification. Those are exactly the data points an age assurance regime would be judged on, and none is promised for the Community Safety Report.

OpenAI is not alone in leaning on inference. Meta said on May 5, 2026, that it would use AI analysis of visual cues such as height and bone structure to find accounts belonging to children under 13, and extended its age prediction to more markets.

Parental controls and violence alerts

Parental controls, which OpenAI introduced on September 29, 2025, let parents link their ChatGPT accounts to their teens' accounts and manage safety settings together. Linked accounts include safety alerts that may be sent when "specially trained reviewers identify a serious self-harm concern and behaviors that may be associated with disordered eating."

In July 2026, according to the post, OpenAI expanded those alerts to notify parents when their teen's account has been deactivated for violating its usage policies on violence. The next step is still in development: OpenAI says it is working on notifications for parents when it pauses a teen's account "for repeatedly discussing violence," linking the alert to the new pause mechanism.

The effectiveness of those controls is disputed in court. Florida's complaint cites an OpenAI LinkedIn advertisement promoting built-in teen safety protections and parental controls, calls that claim false, and alleges that the free tier has no age verification at all. The complaint also includes a count based on the COPPA children's privacy rule. OpenAI has not been found liable on any of those claims.

What the announcement does not settle

Beyond the missing figures, the rollout itself is loosely defined: "the next few weeks," with no dates by feature and no list of markets. There is also the question of where the lines sit between the three tiers the post creates. All users face new limits on reconstructions of past attacks. Users whose recent chats show warning signs get more conservative answers. Predicted minors get those conservative answers regardless, plus a separate set of weapons and security system restrictions. How the tiers interact, for example when an adult is misclassified as a teen and then triggers a warning-sign rule, is not described.

Why this matters for the marketing community

The post does not mention advertising. Its mechanics, however, touch the advertising business that OpenAI has been building inside ChatGPT since January.

When OpenAI announced its first ad tests on January 16, 2026, it said it would exclude advertising from accounts where users indicate they are under 18 or where it predicts underage usage, and would keep ads away from sensitive or regulated topics such as health, mental health and politics. By October 21, 2026, OpenAI said ChatGPT ads were available in 89 countries and that more than 100,000 advertisers had run campaigns, on the Free and Go plans only. The age prediction that decides which safety rules apply is therefore also, by OpenAI's own January description, a gate on which accounts can be shown ads. A model that now reassesses users continuously, and moves accounts into the under-18 experience as new signals appear, can in principle change an account's ad eligibility mid-life. That is PPC Land's reading of the two policies together; OpenAI has not described the interaction, and the post gives no share of accounts it classifies as minors.

Adjacency is the second link. ChatGPT ads appear beneath answers when a sponsored product matches the conversation, so the content of the answer is the environment the ad sits in. OpenAI started brand suitability evaluation pilots with DoubleVerify and Integral Ad Science on October 5, 2026, under which the partners assess how OpenAI's brand safety standards are applied without accessing private user conversations. Neither the October post nor the pilot announcements say how conversations that ChatGPT ends, or answers in its more conservative mode, are treated for ad serving.

The regulatory backdrop is also converging on age. On October 28, 2026, two days before OpenAI's post, New York Governor Kathy Hochul signed S9051-B, which bars operators from offering minors 12 unsafe AI companion features unless they establish that a user is an adult, with penalties of up to $25,000 per violation from January 1, 2027. Its definition of a covered AI companion is not limited to the companion chatbot category: Common Sense Media, which helped draft the bill, counts ChatGPT, Gemini and Claude as in scope. California's SB 1119, signed on September 10, 2026, requires AI companion operators to assess risks and submit findings to an independent auditor. The Federal Trade Commission's September 10, 2025 orders to seven AI companies, OpenAI among them, demanded revenue figures, advertising included, broken down by user age group. And when Character.ai entered Chapter 11, attorneys general for 39 states and the District of Columbia set conditions for any buyer that include age assurance that does not rely solely on self-declared birthdates and account-level advertising opt-outs.

Against that list, the announcement puts OpenAI's own mechanisms on record while regulators and courts are still defining theirs. Whether the first Community Safety Report contains numbers that advertisers, parents and lawmakers can check will be the first measurable test of those commitments, and it is not due before early 2027.

Timeline

Summary

Who: OpenAI, the developer of ChatGPT, together with ChatGPT users of all ages, users it predicts are under 18, their parents, law enforcement agencies that receive OpenAI's referrals, and Persona, the company's age verification partner.

What: OpenAI disclosed more than 200 law enforcement referrals in the third quarter of 2026 and committed to a twice-yearly Community Safety Report. It announced new limits on reconstructions of past mass attacks for all users, more conservative answers after warning signs in recent chats, automatic chat-ending and possible account pauses for users fixated on violence, restrictions on practical weapons and security system information for teens even without violent intent, continuous age prediction, and planned parental notifications for paused teen accounts.

When: The announcement is dated October 30, 2026. The safeguards begin rolling out in the next few weeks. The first Community Safety Report is due in early 2027, covering July to December 2026, with updates every six months.

Where: ChatGPT, with no markets or regional exceptions specified in the announcement. The referral example OpenAI links to involves a case in Brazil passed on through the FBI.

Why: OpenAI presents the changes as part of strengthening its violence safeguards and being transparent about how well they work. They arrive as state laws in New York and California, a Florida lawsuit, private suits and federal inquiries focus on how AI chatbots treat minors and handle violent intent, and as age prediction also decides which ChatGPT accounts are eligible to see ads.