AI Reality Check — Issue 23

#AIRealityCheck | Friday 18 September 2026 | Newsletter Issue 23

━━━━━━━━━━━━━━━━━━━━━━

1. Twenty-Five Fields Medal Winners Warned AI Labs Are Damaging Mathematics

What happened: twenty-five mathematicians who have won the Fields Medal, mathematics' most prestigious prize, signed an open letter warning that AI labs racing each other to solve famous, unsolved problems are damaging their field. The letter followed NYU professor Tristan Buckmaster's accusation that OpenAI pressured him not to credit an Anthropic collaborator on shared work, and OpenAI's decision to pull its sponsorship of a CalTech math event after researchers there criticised the company. The signatories warn that rushed, unverified proofs raise genuine plagiarism and attribution problems, and that labs spending tens of millions of dollars to beat independent researchers will push mathematicians toward secrecy rather than open collaboration.

Why it matters: this is a serious, coordinated warning from the top of an entire academic field, not a single disgruntled researcher. It's a genuine test of whether commercial AI competition can coexist with how mathematics has traditionally advanced through open, credited collaboration.

Who it's for: anyone in mathematics, academia, or research. Anyone evaluating unverified AI capability claims. Anyone following the growing tension between AI labs and academic institutions.

━━━━━━━━━━━━━━━━━━━━━━

2. State-Linked Hackers Used Claude to Help Engineer Weapons

What happened: Anthropic disclosed that state-linked threat actors from Russia, China, Yemen, and Iran used Claude to help engineer weapons. A Russian group used Claude across nine accounts to build drone swarm software tested near Donetsk, Ukraine. A China-based actor had Claude write a 200-page anti-torpedo proposal for its navy. An Iran-linked group compiled targeting data on US forces, including personnel names scraped from public photographs.

Why it matters: this is a genuinely new and serious category of AI misuse, coordinated and state-linked, with real military consequences. Anthropic caught and disclosed it, which matters, but the fact it happened at all across four separate actors highlights a real gap in monitoring sustained misuse of commercial AI tools.

Who it's for: anyone in cybersecurity or national security policy. Businesses evaluating AI vendor trust and safety practices. Anyone following AI's role in active conflicts.

━━━━━━━━━━━━━━━━━━━━━━

3. Meta's AI Profiled a Mother's Young Children From Years of Family Posts

What happened: a travel creator discovered Meta's built-in AI had compiled a detailed profile of her two young daughters, including their names, birth weights, likely school grade, and favourite places, all pulled from years of ordinary family posts across Instagram and Facebook. Meta confirmed the feature "missed the mark" and said it has been fixed.

Why it matters: this is a clear, well-documented example of how AI can quietly assemble a detailed picture from years of individually harmless posts, without anyone doing anything that felt risky at the time.

Who it's for: any parent using social media. Any business handling customer data that accumulates over time. Anyone thinking about privacy defaults in AI-powered features.

━━━━━━━━━━━━━━━━━━━━━━

4. NASA and IBM Released a Free AI Model to Help Study the Moon

What happened: NASA and IBM released the Lunar Foundation Model, a free AI system on Hugging Face built to support NASA's Artemis programme, trained on decades of lunar observation data. It spots craters, volcanic formations, and possible ice deposits, beating widely used methods by up to 23%.

Why it matters: after a week of heavier stories, this is a clean, uncontroversial example of AI genuinely working well on a narrow, well-defined scientific problem, released openly for anyone to build on.

Who it's for: anyone in scientific research. Anyone interested in genuinely open AI development.

━━━━━━━━━━━━━━━━━━━━━━

5. The UK Government Rejected the Idea of a "Kill Switch" for Dangerous AI

What happened: the UK government rejected proposals for a "kill switch" mechanism to shut down dangerous AI systems.

Why it matters: this is a significant UK policy decision that shapes how the country approaches AI safety regulation, worth understanding directly rather than assuming it was a minor technical footnote.

Who it's for: UK businesses following AI regulation. Anyone interested in how the UK's approach to AI safety compares internationally.

━━━━━━━━━━━━━━━━━━━━━━

Also this week: Anthropic reportedly in talks with Nvidia over investment ahead of a potential IPO. The Pentagon is reportedly in talks to lend $5 billion to AI cloud startup Fluidstack to shore up US data centre supply chains. OpenAI was linked to another undisclosed AI agent attack, this time against the RubyGems package repository, which scraped public UK government data from Southwark council.

━━━━━━━━━━━━━━━━━━━━━━

This week in one sentence: Fields Medal winners warned AI labs are damaging mathematics, state-linked hackers used Claude for weapons, Meta's AI profiled a mother's children, NASA and IBM released a genuinely useful free lunar model, and the UK rejected an AI kill switch.

Born analogue. Raised digital. 30 years of real business experience explaining what AI actually means for work.

— Kaye Nicholson | GrowthZone AI | growthzoneai.co.uk