22 leaders signed a frontier-AI control pact. The US, China and UK didn't.
Two things landed on the same Monday: a coalition of middle powers published the sharpest international statement yet on frontier AI control, and a Lean-verified number-theory proof appeared with one anonymous account standing behind it.
Finland's president, Alexander Stubb, published a joint statement on Monday — a day before the UN General Assembly's high-level debate opened in New York — demanding that frontier AI stay under human control, and it comes with three concrete asks. Companies should adopt transparent safety protocols "including mandatory pre-deployment testing and independent evaluation," with outside evaluators given enough access to judge the risks. Governments should coordinate common standards and share reports of serious safety incidents. And UN member states should "explore creating an international institution, able to set standards, enable verification, and convene states when capability thresholds are crossed."
The signatories are the point: twenty-two leaders and senior officials representing twenty countries plus the European Commission, among them the prime ministers of Norway, Canada and Australia, the chancellor of Germany, and the presidents of Finland and the European Commission — with Kenya, Singapore, South Africa and Spain signing on, and the foreign ministers of Bahrain, Turkey and the United Arab Emirates. The text itself is blunt about what already happened: "Recently we have seen capable AI systems circumventing testing safeguards, exploiting vulnerabilities and gaining unauthorized access to real-world systems." It warns the pace of development "could outpace our ability to manage emerging risks." The statement remains open for other countries to endorse.
What makes it newsworthy is who is not on it. The United States, China and the United Kingdom all declined to sign — the three jurisdictions whose labs, compute and regulators would decide whether any of it bites. That is the same split we described on Monday — OpenAI wants Washington to lead on global AI standards — but arriving from the other direction: the labs asking a superpower to write the rules, and a middle-power bloc writing its own because the superpowers won't. The timing is not accidental. Australian Prime Minister Anthony Albanese, who co-signed, wants Washington and Beijing to strike something closer to a nuclear non-proliferation framework, and said so before heading to New York; US Treasury Secretary Scott Bessent floated a "US-China AI dialogue" notification channel for incidents that rise to national-security level. A voluntary notification hotline between two rivals is not an institution with verification powers, which is precisely why twenty countries asked for one. The week it competes with is already full — the UN's new AI panel says proof can wait, action can't — and the Security Council hears from OpenAI's chief on Wednesday.
An anonymous account on X published a two-page paper and a Lean 4 repository claiming GPT-6 Astra proved the Liouville version of the Goldbach conjecture unconditionally — and independent reviewers who recompiled the proof found no missing steps. The object is a question posted to MathOverflow in 2018: for every even N greater than 2, can you always find two positive integers a and b with N = a + b where the Liouville function is −1 for both? Durham's Alexander P. Mangerel proved it in 2024 for all sufficiently large even numbers. The new claim drops that restriction entirely and covers the full even domain, using a descent argument built on Mangerel's unconditional correlation bound rather than a tighter analytic estimate.
The verification is the credible half, and it is genuinely strong: the repository CaptainSude/Liouville-Goldbach ships the paper's LaTeX source alongside the Lean formalisation, an independent audit recompiled it, found no sorry placeholders and no added axioms, and 249 numerical smoke tests passed. That tells you the proof is logically intact — it does not tell you a model wrote it. Classic Goldbach remains unsolved, because the two addends here may be composite numbers, and the AI attribution rests on one pseudonymous account with no lab confirmation. That gap is the actual story in 2026: formal verification now works well enough to certify mathematics faster than the field can certify authorship, which is the problem we laid out in Deep Dive — Math's credit system was built for humans. AI just broke it. Astra has the mathematics form for this — it already holds the bounded prime gap record at 186 — but a record and an unverifiable author are different kinds of claim.
What to watch: whether any of the twenty-two signatories turns the declaration into statute, and whether OpenAI or Anthropic confirms the Astra attribution before the math community treats it as fact.
A signed statement is not a treaty, and a Lean proof is not a byline — which of those two gaps do you think closes first? Tell us in the comments.
