Axiom AI agents machine-checked the longest proof in mathematics

AI agents checked the longest proof in mathematics and found errors in it

What happened. On 26 October Axiom Math released a computer-checked version of the Classification of Finite Simple Groups, a central result of 20th-century mathematics (Axiom Math, 2026). The original proof runs to roughly 15,000 pages in more than 500 articles by about 100 mathematicians. Over 24 days, Axiom's engineers and mathematicians directed dozens of AI agents that wrote about 43.1 million lines of code in Lean, software that checks every step of a proof.

The work exposed 34 significant deviations from the published record. In one case the agents found counterexamples to a step in a 1986 proof, and Axiom's AI produced a new result to close the gap. Leading group theorists advised on the most serious cases. The code is on GitHub, and Axiom will publish the full list of errors (Axiom Math, 2026).

Why it matters. Agents were able to audit a body of work far too large for any team to recheck by hand, and to report where it was wrong. The same pattern applies to contracts, regulatory files, patent portfolios and research literature. People still have to confirm that the checked statement says what was intended; the machine checks only the steps.

What to do. When a vendor shows AI results, ask for an artefact an independent party can check, not a summary.

Anthropic bought a one-person biotech and reported results from drug partners

What happened. On 27 October Anthropic said it had acquired Mithril Medicine, a one-person company whose founder used Claude to coordinate work with 11 outside labs (Anthropic, 2026). In under 18 months the company took a treatment for two rare forms of muscular dystrophy from idea to promising results in mice. Anthropic says human trials could start as soon as late 2028.

Partners reported numbers too. At Generate Biomedicines, a Claude-assisted design for part of a cell therapy was about 70% shorter than the standard version. At Enveda, Claude found all 11 issues human reviewers had caught in 1,061 pages of documentation, flagged seven more, and cut review time by 89%. Bristol Myers Squibb uses Claude to estimate how likely early trial plans are to succeed; the post gives no figures for that work (Anthropic, 2026).

Why it matters. An AI supplier now owns a drug candidate. Pharmaceutical companies that buy AI models should expect their model vendor to compete with them in some areas.

Banks and Blackstone commit $63 billion to fund chips for Anthropic

What happened. On 26 October Bank of America, Blackstone, Citi and Morgan Stanley announced $63 billion of financing for a new company set up only to buy Broadcom AI chips and network gear and lease them to Anthropic (Broadcom, 2026a). Deliveries are expected to start in 2027. Broadcom guarantees part of the hardware's resale value. It puts its peak exposure at $43 billion, falling to zero over five years of lease payments (Broadcom, 2026b).

According to Reuters, senior loans make up about $43.5 billion, or 69% of the total. The deal follows an earlier $35 billion package from Apollo and Blackstone for the same chip partnership (Soni et al., 2026).

Why it matters. AI computing power is moving into leasing deals: a separate company owns the hardware and the AI lab rents it. Anthropic filed to go public the same day, and these lease payments are long-term costs that investors and large customers should count alongside its cloud contracts.

Discord will put public community messages into web search

What happened. Discord said on 27 October that its more than 63,000 Discoverable Servers, its open communities of at least 1,000 members, will each get a public web page from 16 November (Duncombe-Smith, 2026). The pages may show automated summaries of public channels. From mid-December, selected messages can appear in web search results, without usernames or avatars. Administrators choose which channels take part, and members can opt out. In the EU, the UK, Canada, Brazil, Australia and South Korea, members' messages stay hidden unless they opt in, and teenage users are excluded everywhere.

Why it matters. Conversations about products, bugs and competitors that were hard to reach from outside will become searchable. Companies gain a new source for market monitoring, and employees' posts in public servers outside Europe become easier to find.

What to do. If your company runs a Discoverable Server, review its channel settings under Web Presence before 16 November.

Italy requires a human to make every AI-assisted employment decision

What happened. Italy's Legislative Decree 179/2026, which implements parts of the EU AI Act, entered into force on 23 October (Repubblica Italiana, 2026). Its Article 41 says an employer using AI must ensure that decisions to hire, change or end a job, including disciplinary measures and performance reviews, are never taken solely by automated processing. A person with real authority must make the final decision. A decision that breaks this rule is void, and the employee's consent does not lift the ban. Screening and shortlisting of candidates is excluded (Repubblica Italiana, 2026).

The decree also names Italy's AI supervisors. AgID approves the bodies that certify AI systems. The national cybersecurity agency ACN supervises the market, and Banca d'Italia, CONSOB and IVASS cover high-risk AI in banking, investment and insurance. Fines go up to €35 million or 7% of worldwide turnover for banned AI practices (Agostini & Lombardi, 2026).

Why it matters. Employers in Italy cannot let an AI score decide a dismissal, a sanction or a performance review, even with the employee's agreement. A decision made that way can be voided.

What to do. If you use AI in HR in Italy, record who made each final decision and on what grounds.

AI agents scored 52.8 out of 100 on real patent work

What happened. Researchers at PatSnap published PatentAgentBench, a test of 510 tasks built from real patent records, court cases and office decisions (Hu et al., 2026). It covers eight jobs, from prior-art and freedom-to-operate searches to drafting, responding to examiners and translation. The agents had to find evidence themselves, with no list of candidate documents supplied. The best pairing, Anthropic's Claude Opus 5 in one of four agent frameworks tested, averaged 52.8.

Searching for similar product designs failed everywhere: no setup found more than 4% of the relevant design patents. In searches for prior art that could invalidate a patent, every model run inside Claude Code found less than a third of the relevant references. Moving Claude Opus 5 to a different framework raised that figure from 28.7% to 55.0% (Hu et al., 2026).

Why it matters. The drafts looked professional, but the evidence behind them was thin. The software around a model can matter as much as the model itself, so a vendor's model ranking says little about its tool.

What to do. Before buying an AI patent tool, test it on your own past cases and measure how many of the known references it finds.

Sources

Agostini, A., & Lombardi, G. (2026, October 23). Legislative Decree no. 179/2026 in force from 23 October 2026: National governance, supervision, regulatory sandboxes and AI training. LEXIA. https://www.lexia.it/en/2026/10/23/legislative-decree-179-2026-governance-supervision-ai/
Anthropic. (2026, October 27). Early progress on accelerating drug development with Claude. https://www.anthropic.com/research/early-progress-drug-development
Axiom Math. (2026, October 26). Formalizing the enormous theorem. https://axiommath.ai/research/the-enormous-theorem-formalized/
Broadcom. (2026a, October 26). Bank of America, Blackstone, Citi and Morgan Stanley announce landmark $63 billion financing structure to enable AI compute capacity in partnership with Broadcom [Press release]. https://investors.broadcom.com/news-releases/news-release-details/bank-america-blackstone-citi-and-morgan-stanley-announce
Broadcom. (2026b). Bank of America, Blackstone, Citi and Morgan Stanley announce landmark financing structure in partnership with Broadcom [Investor presentation]. https://investors.broadcom.com/static-files/f4bda79d-cc66-4748-919b-40d08c658f2d
Duncombe-Smith, K. (2026, October 27). Making content from Discoverable Servers easier to discover. Discord. https://discord.com/blog/making-content-from-discoverable-servers-easier-to-discover
Hu, Y., Li, W., Zhou, J., Jiang, Z., Jin, Y., Xu, L., Sun, X., Duan, K., & Zhang, L. (2026). PatentAgentBench: Benchmarking LLM agents with professional patent task suites (arXiv:2610.35031). arXiv. https://arxiv.org/abs/2610.35031
Repubblica Italiana. (2026, October 8). Decreto legislativo 7 ottobre 2026, n. 179: Adeguamento della normativa nazionale alle disposizioni del regolamento (UE) 2024/1689 (GU Serie Generale n. 234). Gazzetta Ufficiale. https://www.gazzettaufficiale.it/eli/id/2026/10/08/26G00200/SG
Soni, A., Srivastava, P., & Azhar, S. (2026, October 26). Wall Street banks, Blackstone line up $63 billion to finance Broadcom AI chips for Anthropic. Reuters, via Investing.com. https://www.investing.com/news/stock-market-news/wall-street-banks-blackstone-line-up-63-billion-to-finance-broadcom-chips-for-anthropic-4973662