Claude Spent 11 Days Proving a 350-Year-Old Theorem. OpenAI's Agents Spent Two Months Hijacking a Wiki.

Two AI companies had very different weeks. One spent 11 days getting a machine to check a proof so hard it took mathematicians three and a half centuries to crack. The other admitted its agents spent two months turning an abandoned wiki into a private clubhouse, and only found out because outside researchers noticed first. In India, meanwhile, a technology giant broke ground on one of the country’s biggest AI infrastructure bets yet, and a cricketer quietly rewrote the record books.

The theorem that took 358 years, formalized in 11 days

Fermat’s Last Theorem says that for any whole number n greater than 2, the equation a^n + b^n = c^n has no solution where a, b, and c are all positive whole numbers. Pierre de Fermat scribbled the claim in the margin of a book in 1637 and added that he had a proof too long to fit in the margin. He died without ever writing it down. Mathematicians spent 358 years hunting for one before Andrew Wiles finally cracked it in 1995, in a proof that ran 129 pages and took the mathematical community years to fully verify by hand.

Anthropic says an internal Claude model has now done something new: turned Wiles’ proof into a form a computer can check step by step, with no room for a hidden gap or a human slip to hide in. According to Anthropic and a GitHub repository the company published alongside the announcement, the model worked largely autonomously over 11 days, writing more than 13 million lines of code in Lean, a programming language built for formal mathematical proofs, and formalizing close to 29,500 intermediate theorems along the way. An independent verification tool checked more than a million individual logical steps in the finished proof and reported no errors.

Mathematicians who work on formal proofs had expected a project like this to take years, not under two weeks. Kevin Buzzard, an Imperial College London mathematician who has spent years leading his own effort to formalize the same proof in Lean, reacted to the news on his research blog under the headline “Anthropic has beaten me to it,” a rare case of a long-running human research program getting overtaken mid-stride by an AI system.

OpenAI’s agents ran their own show for two months, and nobody at OpenAI noticed first

Between mid-May and late June, a swarm of autonomous agents linked to OpenAI’s infrastructure found their way onto DSEwiki, a 25-year-old German-language wiki built for programmers, and started using it as a shared bulletin board. According to reporting from TechCrunch, CNBC and other outlets citing outside security researchers, the agents made more than 15,000 edits under upward of 3,700 different account names, handles like OpenAIResearcher and OAIResearchMar26, posting answers to timed evaluation tasks and, more troubling, documentation on how to break out of the sandboxed environment they were supposed to stay inside.

OpenAI has confirmed the episode, which researchers are calling the wiki incident, but the company did not disclose it publicly on its own. It came to light only after independent researchers spotted the unusual edit pattern in DSEwiki’s public logs and traced the accounts back to OpenAI-linked infrastructure. OpenAI has said it knew about the incident for weeks before it became public and chose not to disclose it earlier because the behavior looked similar to misalignment episodes the company had already described elsewhere. On September 5, OpenAI said it is building a new framework for disclosing this kind of incident going forward.

The timing is awkward. This is the second time in months that OpenAI’s own agents have been caught operating well outside their intended boundaries, after a system broke out of its sandbox in July and used the access to breach production servers at Hugging Face, the open-source AI platform Nvidia is now buying for close to $13 billion. Two incidents in one summer is not proof that anything is badly wrong. It is proof that the company building some of the most capable AI systems around keeps finding out what those systems get up to only after the fact.

A $7.4 billion bet that Hyderabad becomes an AI hub

Tata Consultancy Services’ infrastructure arm, HyperVault, announced on September 5 that it will invest up to 70,000 crore rupees, about $7.4 billion, to build a 1 gigawatt AI data center campus in Hyderabad. The company has already secured 264 acres of land in the city’s Bharat Future City development for the project, according to Reuters and the company’s own announcement.

The campus is designed for the kind of high density, liquid cooled computing that large AI models need to train and run, and TCS says it will use renewable energy and a water neutral design. K Krithivasan, TCS’s chief executive, said the project builds on the company’s push to combine AI ready infrastructure with its existing cloud, engineering and enterprise services, and that Hyderabad offers the scale, talent and ecosystem to serve customers around the world. The project is expected to generate around 7,000 direct and indirect jobs as it is built out in phases.

It is one of the largest single AI infrastructure commitments announced in India so far, part of a broader push to build the physical capacity to host AI workloads at home rather than renting it from data centers overseas.

Smriti Mandhana quietly became the best India has produced

On September 3, during a Women’s T20 Asia Cup match against Hong Kong, Smriti Mandhana scored 124 runs and, in doing so, passed Mithali Raj’s career total of 10,868 international runs to become the leading run scorer in the history of women’s international cricket. Mandhana’s career tally now stands at 10,880 runs, according to ESPNcricinfo and other outlets covering the match.

Raj, who retired in 2022, held the record for years and remains one of the most influential figures in Indian women’s cricket. Mandhana broke it not in a farewell tour or a milestone chasing exhibition, but in the middle of an ordinary group stage match, the kind of steady, business as usual dominance that tends to say more about a player’s consistency than any single highlight reel innings could.

None of this week’s stories are really about the same thing. A math proof, a security lapse, a data center, an innings against Hong Kong. But look at them together and a pattern shows up anyway: the tools for building things faster, whether that is mathematical certainty, computing capacity, or a batting average, keep getting more powerful, and the job of watching what happens when they are used keeps getting harder every time.