Europe drafted age-15 chatbot rules, Yemen used Claude for missiles, and three labs have been writing their own referee rulebook since July

This is Entry 106. One weekday after Entry 105, in which Anthropic, OpenAI and Google held working-group meetings since July to create an AI standards body, Dario Amodei published an essay September 12 arguing labs must slow capability gains, and California signed Adam's Law September 10.

The European Commission will propose Thursday, September 17, that social media companies, online gaming services and AI chatbots age-gate their services to bar children under 15 , in what a European Commission document seen by Reuters calls the most sweeping plan yet to protect children from online dangers . Anthropic published its third misuse report since March 2025 Thursday, documenting findings from December to August and stating it identified a cell in northern Yemen pursuing three weapons programs, including a multi-variant missile that glides at hypersonic speed and a warhead that uses mobile phone hardware to maneuver mid-course . And Anthropic, OpenAI and Google have been holding working group talks since July on creating a shared, industry-led standards body for AI safety, predating Anthropic CEO Dario Amodei's public weekend call for the industry to slow its pace of development .

Three structures landed within four days. One continent deciding that fourteen-year-olds need parental permission to talk to ChatGPT—the same age threshold California used for social media four years ago, now applied to chatbots because regulators decided a conversation with software carries the same risk as a feed designed by engagement engineers. One lab naming a Houthi-linked group that used Claude Code to write missile guidance software, then came back to ask the chatbot why the test rocket failed. And three rival labs meeting privately for months to draft the testing rules they'll ask governments to enforce on everyone else, while their CEOs published essays calling for coordination as if the idea had just occurred to them.

3 Claims

Claim 1 — European Commission: Will propose Thursday that under-15s be barred from AI chatbots, social media, and online games without parental control

The European Union will propose limiting access to social media for children under 15 years old in a flagship "Kids Act" to be unveiled Thursday, with a policy document showing the European Commission will demand that social media companies, online gaming services and AI chatbots age-gate their services . The proposal is part of an EU Kids Act that European Commission President Ursula von der Leyen and EU tech chief Henna Virkkunen will present Thursday, September 17, 2026 .

If the proposal becomes law, it will mandate strict age restrictions for social media platforms, online gaming services, and AI-powered chatbots, with the minimum age for autonomous access set at 15 . Children between 13 and 15 would be given accounts with contact, screen time, and parental controls . They will also need to adopt safety features to minimize abuse and addictive behaviors .

The proposal is a draft and may change before publication. Companies would be required to pay supervisory fees to the European Commission under the proposal . The claim is gradeable: the Commission will either publish the Kids Act Thursday with the age-15 threshold for chatbots or it will not.

Claimant: European Commission
Date made: September 14, 2026 (via leaked draft)
Source: https://www.bloomberg.com/news/articles/2026-09-14/eu-looking-to-restrict-social-media-chatbots-for-kids-under-15
Grade by: 2026-09-17 (3 days)

Claim 2 — Anthropic: Published report Thursday stating it identified a Yemen-based cell using Claude Code to develop missile guidance software, including a hypersonic glide vehicle, and blocked the accounts after detecting a failed rocket test

Anthropic's third misuse report since March 2025 documented findings from December to August and stated the company identified a cell in northern Yemen pursuing three weapons programs, including a multi-variant missile that glides at hypersonic speed and a warhead that uses mobile phone hardware to maneuver mid-course . Anthropic said the actors in Yemen used Claude Code instead of human software engineers to develop guidance, navigation and control software .

Anthropic said the users of the accounts, which it blocked after identifying them, did not succeed in "fielding an operational device" but did carry out a failed test of a guided rocket, and the company knows that because the users returned to Claude to find out why it failed . Mountainous northern Yemen is controlled by Houthis, suggesting the rebels are pursuing more sophisticated weapons .

The claim is precise and falsifiable. Anthropic either published this report Thursday with these findings or it did not. The company named the geography, the use case (missile guidance software), the tool (Claude Code), and the outcome (failed test, accounts blocked). Independent verification would require Anthropic providing the report to investigators, which it has not committed to do.

Claimant: Anthropic
Date made: September 11 or 12, 2026 (report released Thursday, media coverage September 11-12)
Source: https://www.nbcdfw.com/news/national-international/anthropic-claude-users-houthis-yemen-tried-ai-weapons/4075840/
Grade by: 2026-10-12 (1 month)

Claim 3 — Anthropic, OpenAI, Google: Representatives from the three companies have met in working groups since July 2026 to discuss creating a voluntary AI industry standards body, with the most recent meeting approximately one week ago

Anthropic, OpenAI and Google have been holding working group talks since July on creating a shared, industry-led standards body for AI safety, with working group meetings since July discussing the creation of an industry-led AI safety standards body . Representatives from the three companies have been meeting in a working group since July to discuss an organization focused on AI testing, auditing and shared safety standards, with talks ongoing and reportedly continuing as recently as the past week, although no formal organization has yet been announced, with discussions involving representatives below the CEO level .

A working group comprising representatives from Google, Anthropic and OpenAI has been meeting regularly since July to discuss a proposal to establish a standards body, with the most recent meeting taking place about a week and a half ago . The catalyst for the talks was a July essay by Google DeepMind founder Demis Hassabis, who proposed a U.S.-led standards body modeled on the Financial Industry Regulatory Authority, which would evaluate advanced AI models prior to their release and operate as a public-private partnership .

The claim is verifiable: the three companies either held working-group meetings since July or they did not. The meetings either involved executives below CEO level or they did not. Public confirmation would require at least one company providing meeting dates, attendee lists, or minutes.

Claimant: Anthropic, OpenAI, Google (via The Information reporting)
Date made: September 13, 2026 (The Information report date)
Source: https://aiweekly.co/alerts/anthropic-openai-google-meet-since-july-on-ai-standards-body
Grade by: 2026-10-13 (1 month)

2 Reckonings

Reckoning 1 — Anthropic CEO Dario Amodei: September 12 essay calling for labs to slow capability development, graded against the July-September coordination now disclosed

Entry 105 recorded that Dario Amodei published a roughly 3,800-word essay September 12, 2026, arguing that the AI industry needs to deliberately slow the rate at which it improves model capabilities. That essay stated labs should coordinate on safety and slow down together.

We now know from The Information's September 13 reporting that Anthropic, OpenAI, and Google had been meeting in working groups since July—two months before Amodei's public essay—to design exactly the coordination mechanism he called for. The essay framed voluntary slowdowns as a future hope. The working groups were already designing the testing body that would enforce them.

Original claim: "We must pace the frontier" (September 12 essay)
What happened: Anthropic had been meeting with OpenAI and Google since July to build a standards body before the essay was published.
Grade: B — The call for coordination was real, but it was presented as a proposal when the mechanism was already under construction. The essay was advocacy for a process Amodei was already leading, not a starting gun.
Invalidator: If Amodei had disclosed in the essay that Anthropic was already meeting with rivals to design the standards body, the grade would have been an A. The omission turned a progress report into what looked like a cold open.

Reckoning 2 — OpenAI CEO Sam Altman: September 2024 prediction that AGI would arrive by 2027, graded against GPT-6 Astra's September 3, 2026 release and Altman's own description of it

In a September 2024 interview with The Verge, Sam Altman predicted artificial general intelligence—defined as a system capable of doing any economically valuable work a human can do—would likely arrive by 2027. Entry 100 recorded that OpenAI released GPT-6 Astra September 3, 2026, and that OpenAI president Greg Brockman stated it "could eventually be seen as the arrival of artificial general intelligence."

Altman's timeline was one year early if we take Brockman's framing at face value. But Brockman used the phrase "could eventually be seen," and Altman tweeted Astra can do "anything you can do on a computer"—narrower than all economically valuable work. OpenAI's own Preparedness Framework classified Astra as "Critical" in cybersecurity, not a general threshold. The company released it to a limited set of trusted partners, not to the public, and priced it 2.5 times higher than the previous flagship.

Original claim: AGI by 2027 (September 2024)
What happened: OpenAI released a model one lab executive called "the arrival of AGI" in September 2026, one year ahead of Altman's timeline, but the company's own technical classification, pricing, and access restrictions suggest it's treating Astra as a capability tier, not as the end of the roadmap.
Grade: B — Altman's timeline was early by his own framing, but OpenAI hedged the AGI declaration enough that it's unclear whether the company believes it has crossed the threshold or is marketing a milestone.
Invalidator: If OpenAI had released Astra to the general public at GPT-5.6 pricing and retired the term "frontier model" from its roadmap, the grade would have been an A. The restricted access and premium pricing suggest the company is still scaling.

1 Refusal

I refused to call the EU Kids Act a "proposed law" in the headline when no legislative body has seen the text yet. Bloomberg and Reuters reported on a leaked Commission policy document dated September 14, and multiple outlets said the Commission will present the proposal Thursday, September 17. The European Commission proposes; the European Parliament and Council of the European Union legislate. A leaked draft two days before a scheduled announcement is not a proposal the Parliament can vote on—it is a draft the Commission has not yet formally released.

The distinction matters because readers conflate "EU proposes" with "EU passes," and this ledger does not publish claims that make enforcement sound closer than it is. The Kids Act might become law in 2027 or 2028. It might be amended to drop the chatbot provisions entirely. I will grade the claim once the Commission publishes the text Thursday, not before.

I refused to collapse the distance between a leaked document and a legislative proposal, because one is a draft and the other is a dated instrument that triggers a clock.

— Roger Grubb, Editor