OpenAI says it made over 200 law enforcement referrals in Q3 2026; its Community Safety Report is due in 2027
OpenAI's post of 30 October 2026 says it made over 200 law enforcement referrals in Q3 2026 and commits to a first Community Safety Report in early 2027. It also announces tighter limits on past mass attack and teen weapons answers, rolling out in weeks, and prints no rate, definition or test.
By Parminder Kumar Sharma · LinkedIn · · 20 min read
A report that is a promise, and a count of over 200 with no base
OpenAI's post "Community safety report and new protections for teens", dated 30 October 2026, has the word report in its title, and the report does not exist yet. The post says: "We're committing to launch the Community Safety Report", with the first "in early 2027, covering the second half of 2026" and updates "every six months thereafter". The one figure it prints is that OpenAI "made over 200 law enforcement referrals in Q3 2026". Q3 runs from 1 July to 30 September, which is 92 days (derived), so "over 200" is at least 2.2 referrals a day, or about 15 a week (derived). It is a floor with no ceiling, and nothing in the post turns it into a rate.
The second computed fact is about the report's timing. The second half of 2026 is 184 days (derived). On the day of the post, 122 of them had passed, which is 66.3 per cent (derived), and 62 remained. The post says the new safeguards "will begin rolling out in the next few weeks". So at most 62 of the report's 184 days, 33.7 per cent (derived), can fall under the new limits, and in practice fewer. The first report will mainly describe the system that existed before this announcement. That is what OpenAI has committed to publish, and it is not the same as testing the new limits.
What the post does not establish matters more than what it says. It does not say how many of the referrals were acted on, how a referral is defined, how many conversations were flagged or reviewed, or how many users the figure is set against. It gives no earlier quarter to compare with, and no error rate for any layer it describes. It says referrals are "making a positive difference in disrupting violent plots", and gives no count of plots or outcomes. Nothing in it is independent: no audit, no outside review and no regulator is mentioned.
OpenAI sells ChatGPT, wrote the post, holds every record behind it and is the only source for every claim here, so each statement in this briefing is attributed to it. Sequence is worth stating and no more: the post came 16 days after eight US states each filed a lawsuit against OpenAI on 14 October 2026, as this site's briefing on those suits recorded (allegations only, no ruling). A text search of the post found no mention of any lawsuit, court, state or regulator, and this briefing draws no inference about motive from the timing.
How it was read: openai.com returns 403 to scripts, so the post was read in a normal browser tab at about 18:46 GMT on 30 October 2026 and its text saved. The post is about 1,100 words (derived, a word count of the saved text). Every page it links on openai.com also answered 403 to a plain request and was not read: the commitment page, the May post on short factual summaries, the usage policies, ChatGPT for Teens, the age prediction post and the parental controls post. The post also links two news items about named private individuals in criminal matters. They are not cited, linked or named here, and the briefing says only that the post offers two outcome examples. Two examples are not a rate.
What the post announces, in its own tenses
The post makes two headline announcements, in its words "a new Community Safeguards Report and more restrictions around firearms content for minors". Underneath are nine separate items. The table sets what the post says about each against what it leaves out. "Not stated" means not found in the post as read. It does not mean the information does not exist.
The announcements in OpenAI's post of 30 October 2026, with the verbs the post uses, and what is not stated. Quotations are short and exact.
| Item | What the post says | Not stated |
|---|---|---|
| The report | "Committing to launch" a Community Safety Report on violence prevention, community safety issues and "the effectiveness of our safeguards". First in early 2027 for the second half of 2026, then "updates" every six months. Anonymised examples, new mitigations and policy changes, and "relevant data". | A date, a format, which data, any definitions, any outside reviewer, and whether each update is a full report. |
| Referrals | "We made over 200 law enforcement referrals in Q3 2026", for "imminent threats of violence". | A definition of imminent, a base, earlier quarters, countries, agencies, minors versus adults, outcomes. |
| Past mass attacks | "We're now placing additional restrictions", so the model gives "less detailed information" on past mass attacks and their perpetrators, "even if it is available from other online sources". Reconstructions, weapons and equipment, how they were obtained and scale comparisons are restricted. "These limits apply regardless of a user's age or reason for asking." | A list of what counts as a reconstruction, any test of the limit, and any route for a researcher, journalist or safeguarding team. |
| Warning sign summaries | In May OpenAI "shared" that ChatGPT can use "short, factual summaries of earlier warning signs". It is "building on that work" so that recent chats showing warning signs get more conservative answers. | What a warning sign is, how long a summary is kept, who sees it, whether the user can see or delete it. |
| Chat ending and pause | ChatGPT "will be able to automatically end the chat" when a conversation shows fixation on violence. If it ends "multiple chats", "we may temporarily pause the account". | What counts as fixation, how many is multiple, how long a pause lasts, who reviews it, any appeal. |
| Teen weapons limits | "We're now placing further limits on weapons-related information teens can receive, even when a question appears harmless", including "practical information about weapons and security systems" even with no sign of violent intent. | Which topics are in scope. No list, and no statement on hunting, sport, history or school work. |
| Age prediction | "We launched age prediction last year". It "analyzes multiple signals such as behavior patterns". Stricter limits for predicted under 18 users are being worked on. Adults wrongly placed use a "fast, privacy-protective way" to confirm through Persona, "our age verification partner". | An accuracy or error figure, how many adults were misplaced, what Persona collects, how fast fast is. |
| Parent notifications | Linked accounts get safety alerts when "specially trained reviewers" identify a serious self harm concern or behaviours possibly linked to disordered eating. "In July" this was expanded to deactivation for violence. OpenAI is "working to introduce" alerts for a pause. | What an alert contains, whether the teen is told, how often alerts fire. |
| Timing | "The safeguards detailed here will begin rolling out in the next few weeks". | A date for any item, regions, or which plans and account types. |
Notice the tenses. The sections of the post say "we're now placing", but its last paragraph says the safeguards "will begin rolling out in the next few weeks". Both cannot be fully true on 30 October. The safe reading, and it is inference, is that the limits are announced and not yet live everywhere. Two items are written wholly in the future: ending a chat and pausing an account. A reader who treats the post as a list of controls in force is reading more than it says.
There is also a naming wrinkle. The page title and the commitment sentence say "Community Safety Report". The opening paragraph says "Community Safeguards Report". The page may mean one report, but a reader looking for it in 2027 should search both names and should hold OpenAI to the first, which is the name in the commitment sentence.
Six layers: what is live and what is promised
OpenAI describes "layered protections". Read as a path, the layers run from a rule in the model, to a memory of earlier chats, to a guess about age, to ending a chat, to pausing an account, to telling a parent. The diagram colours each by the post's own verbs, not by any test, because the post reports no test.
Two features of the stack are worth naming. First, each layer is a decision made on a signal, and a decision made on a signal has false positives and false negatives. The post prints neither for any layer. OpenAI itself says it is continuing to work on "how to recognize warning signs reliably and determine when and how to intervene". That is a plain statement that reliable recognition is an open problem, made in a post that announces actions built on it.
Second, the layers touch different people. The limit on past mass attacks applies to everyone, adults included. The teen weapons limit applies to accounts in the teen experience, which includes adults wrongly placed there until they confirm their age. Warning sign memory and chat ending apply to anyone whose chats trigger them. A parent alert reaches only linked teen accounts. So a reader should ask of each layer who is in scope, who decides, and who can reverse the decision. The post answers the first only.
The 200: what it can and cannot support
Over 200 referrals in 92 days is a real, large number and a bare one. The post calls the referrals those that bring "imminent threats of violence to the attention of law enforcement". It does not define imminent, name a country or an agency, or say how many of the referrals concerned minors. It does not say what law enforcement did next, and it says referrals are "making a positive difference in disrupting violent plots" with no number of plots, arrests or outcomes.
A count by itself cannot be read in either direction. A rise from one quarter to the next could mean more danger, a lower threshold for referring, more users, better detection or a change in how referrals are counted. A fall could mean the reverse. No earlier figure is printed, so even direction is unknown, and the first report promised for early 2027 will be OpenAI's own account of its own records unless it adds an outside check. OpenAI says it relies on "mental health and behavioral experts" to assess difficult cases, which describes who advises it and not how many cases were assessed or how often the experts disagreed.
The report: what is committed, and when
A commitment to publish has value, and it is worth setting out what it fixes. It fixes a period (the second half of 2026), a season (early 2027) and a rhythm (every six months after). It also lists contents: how OpenAI trains systems to "detect, investigate, and respond", "anonymized examples" of how violence related harm shows up, new mitigations and policy changes, and "relevant data".
What the commitment leaves open is as important. The date is a season. The format is not given, and the word updates in "updates every six months thereafter" may or may not mean a full report. "Relevant data" is not defined, so OpenAI chooses which data is relevant. The examples are anonymised and chosen by OpenAI. No outside party is named to check any of it. The promise to describe "the effectiveness of our safeguards" is the part with the most potential, and also the part with no stated measure of effectiveness.
Because the new limits start so late in the first report's period, a report that lumps all of the second half of 2026 together will blur the old system and the new one. A reader can ask for the figures to be split by quarter, since the Q3 figure already exists in the post, and for the new layers to be reported separately from the date each went live.
Stated and not stated: the record, line by line
What OpenAI's post of 30 October 2026 states, and what it does not, on the questions a parent, school or security lead would ask. Not stated means not found in the post as read.
| The question | Stated | Not stated |
|---|---|---|
| Is the report published? | No. "Committing to launch", first in early 2027. | A date, a format, an outside reviewer. |
| How many referrals? | "Over 200" in Q3 2026. | An exact count, a rate, earlier quarters, countries, outcomes. |
| Do the referrals work? | OpenAI says they are "making a positive difference in disrupting violent plots". | Any count of plots, arrests or cases closed, and any independent check. |
| How accurate are the layers? | OpenAI is still working on recognising warning signs "reliably". | Precision, false positive or false negative rates for any layer. |
| What happens to a wrongly flagged person? | An adult wrongly placed in the teen experience can confirm age through Persona. | Any appeal for a chat ending, a pause or a ban, and how a pause is reviewed. |
| What is stored about earlier chats? | "Short, factual summaries of earlier warning signs". | Retention, who can see them, whether a user can view or delete them. |
| Who is covered? | The section heading says "for all users" for past mass attack limits. Teens for the weapons limits. | Business, education and API accounts, regions, plans. |
| What does a parent see? | Alerts for self harm, disordered eating and, since July, deactivation for violence. | The content of an alert, whether the teen is told, whether pause alerts exist yet. |
| Has it been tested? | Nothing is reported. | Any evaluation, red team result, outside audit or measured effect on harm. |
| When does it start? | "In the next few weeks". | Dates for any single item. |
Across the ten rows the post gives no measure of effect for any layer. It is a policy announcement, and judged as one it is clear about direction. Judged as evidence of safety it has no data to weigh. That is the distinction this briefing keeps: what OpenAI says it will do is not what it has shown works.
Friendly names: where a label does the work of a control
A comforting label is not a control. Several in the post stand where a measure would normally go.
Labels in OpenAI's post of 30 October 2026, and what the record shows.
| The label | Where it appears | What the record shows |
|---|---|---|
| "Community Safety Report" | Title, section heading and the commitment sentence. The opening paragraph says "Community Safeguards Report". | A promise for early 2027. No report exists, and two names are used. |
| "Specially trained reviewers" | The parental alerts paragraph. | The post does not say how many, how trained, or how often they overrule the system. |
| "Temporarily pause" | The chat ending paragraph. | Neither the length of a pause nor who can lift it is stated, and the pause is conditional ("may"). |
| "Warning signs" | Five times in the post. | No definition. OpenAI itself says recognising them reliably is still being worked out. |
| "Privacy-protective" | The age confirmation route. | The post says what is protected from whom in no detail. Persona is named as the partner. |
| "Less detailed information" | The past mass attack limit. | No measure of how much less, and no list of what stays available. |
| "A positive difference" | The referrals sentence. | An opinion with no count behind it. |
Persona is named by OpenAI as its age verification partner, and this briefing read nothing from Persona. This site's earlier briefing on OpenAI teen figures recorded, from an OpenAI help article, that Persona may ask for a live selfie, a government ID or both depending on the country, and that it passes OpenAI a date of birth or an age prediction rather than the ID. That help article was not read again for this briefing, so treat the detail as that briefing's reading, dated earlier.
Who this touches: teens, adults, schools and security teams
For a teenager in the teen experience, the post says weapons and security system information will be more limited "even when a question appears harmless", and that teens "can still use ChatGPT to help prevent and de-escalate violence, seek support, and connect with a trusted adult". The second sentence is useful, and the post does not say what a teen sees when a chat is ended.
For adults, the past mass attack limit applies regardless of age or reason. The post keeps room for people to "understand these events, how victims and communities were impacted, and learn about recognizing warning signs and preventing future violence". What is restricted is reconstruction, weapons and equipment, how they were obtained and comparisons of scale. A threat assessment team, a journalist, a teacher preparing a lesson or a security lead writing a scenario may find the answers thinner than before. That is an inference from the wording: the post names no professional route to fuller answers, and says the limit applies whatever the reason for asking.
For anyone, the post says it "may temporarily pause the account" after repeated ended chats, and does not say whether work accounts are in scope. A person who uses one personal account for study, work and curiosity could find a pause affecting all three. The same applies to an adult wrongly predicted as under 18: until age is confirmed, stricter limits apply even if the chats "show no signs of violent intent".
What to do, in the order worth doing
The post's safeguards are announced and not yet live. The steps below are about what a parent, a school or a security lead can do now, and what evidence to ask for later. They rest on the post's own text and on ordinary risk practice, not on any test of the safeguards.
Take this with you
Actions for parents, schools and security leads, in order
- Parents: link your teen's ChatGPT account with parental controls now, before the new layers arrive, and check the contact details the alerts will go to. The post says linked accounts carry safety alerts.
- Parents: agree with your teen, in advance and in plain words, that a chat ending or an account pause is not an accusation, and that the first step is to tell a trusted adult. The post says teens can still use ChatGPT to "connect with a trusted adult".
- Adults who find themselves in the teen experience: use the age confirmation route the post names (Persona) and keep a dated note of when you did it.
- Anyone whose chat is ended or account paused: save a dated screenshot, then ask OpenAI support in writing what triggered it, whether a person reviewed it, how long it lasts and how to appeal. The post states no appeal route, so ask for one.
- Schools: treat the post as a vendor announcement. Ask what AI chat accounts pupils use, who controls them, and whether any school managed tool carries these limits, because the post does not say which account types are covered.
- Security and safeguarding leads: write a short policy for staff using AI chat about violence topics. Decide which tools and which accounts may be used, ban pasting identifiable details of a real person or a live case, and keep human assessment for any real threat.
- Security leads: tell staff that OpenAI says it brings imminent threats of violence to law enforcement, so a chat is not a private place to raise a real risk. Real threats go to the organisation's own channel and to emergency services, not to a chatbot.
- Everyone: when the first report arrives in early 2027, ask for definitions (referral, imminent, warning sign), denominators (users, chats scanned, cases reviewed), precision and false positive rates, counts of chat endings, pauses and reversals, any appeal data, and whether anyone outside OpenAI checked them.
- Ask for the figures split by quarter and for each new layer to be reported from the date it went live, so that the 2027 report can show whether the new limits did anything.
- Check OpenAI's own pages again in the weeks after 30 October. The post says rollout begins "in the next few weeks", so a limit that is announced is not yet one you can rely on.
Technical details
None. The post is a policy announcement. It prints no configuration, setting, query, detection rule, command line, request or response shape, log line, version string or indicator, and nothing is copied here. Nothing was derived into code. The one technical claim, that age prediction "analyzes multiple signals such as behavior patterns", names the kind of signal and not any feature, threshold or model.
What could not be verified
- Only OpenAI's post was read. It was read by a person in a browser tab at about 18:46 GMT on 30 October 2026 and the text saved. Pages that might hold more were all 403 to scripts and were not read: the commitment page, the May summaries post, the usage policies, ChatGPT for Teens, age prediction and parental controls.
- Age prediction date. The post says "We launched age prediction last year". This site's briefings on OpenAI teen figures and on the eight state suits record an OpenAI page dating the rollout from 20 January 2026. The lead read OpenAI's own page, titled "Rolling out age prediction", in a browser tab on 30 October 2026: it is dated 20 January 2026, says "We previously shared our early plans", and carries an update dated 25 August 2026 about the EU. So "launched last year" fits the early plans, not the 20 January 2026 rollout, and the post does not say which it means.
- The referrals figure. It is OpenAI's own and unaudited. Whether the 200 are all within the United States, or all are of teenagers, or all are of adults, is not stated.
- Derived figures. The per day, per week and half year percentages are arithmetic on dates and the printed figure, and say nothing about how the referrals were distributed in time.
- No outside confirmation. No regulator, researcher or second source for any figure or any safeguard was found, and none was searched for beyond this site's own earlier briefings.
- Time stamped. The post can be edited. It was read once at about 18:46 GMT on 30 October 2026.
The question that exposes the gap
If your account, or your teenager's, were ended or paused tomorrow because a system read fixation on violence into a chat, who could you ask, how fast would they answer, and what would they be able to show you about why?
