Google's 9x Math Triumph & Copilot's Security Hole

Google DeepMind's AlphaProof Nexus solves nine open Erdős problems, topping OpenAI's math breakthrough. Plus: Claude Mythos finds 10,000+ security vulnerabilities, a critical Copilot Cowork exploit, the one-person $250M startup, and Nvidia becomes the first $5 trillion company.

Transcript

SPEAKER_00 0:01

Hey, welcome to AI Mornings with Andreas Vig. It's Monday, the 26th of May 2026. The math wars are heating up. Last week, OpenAI made headlines when its AI cracked an 8-0-year-old AirDose conjecture. Well, Google just said hold my beer. DeepMind's Alpha Proof Nexus has solved nine open AirDose problems, including two that had stumped mathematicians for 56 years. The system pairs a large language model with Lean, a proof assistant, to generate machine-verified mathematical proofs. Each problem costs just a few hundred dollars to solve. The AI also proved 44 open conjectures from the online encyclopedia of integer sequences. Problems requiring entirely new mathematical constructions are still out of reach, but this shows how formal verification is letting AI make genuine mathematical discoveries at machine speed. The rivalry between OpenAI and Google in mathematical reasoning is getting intense. Speaking of AI cracking tough problems, Anthropic just dropped the first results from Project Glasswing. Claude Mythos, their specialized security model, has found over 10,000 high or critical severity vulnerabilities in just one month, working with about 50 security partners. Cloudflare alone found 2,000 bugs with a false positive rate better than human testers. Mozilla fixed 271 vulnerabilities in Firefox. Anthropic also scanned over a thousand open source projects and 62% of the flagged issues held up under independent review. One partner bank even used Mythos to detect and block a $1.5 million fraudulent wire transfer. Anthropic says Mythos remains gated because no company, including itself, has safeguards strong enough to prevent misuse. But with OpenAI ramping up cyber models and Chinese players catching up, equally capable AI will emerge eventually. Here's a security vulnerability you should know about if your company uses Microsoft 365. Researchers at Prompt Armor discovered that Microsoft Copilot Cowork is vulnerable to file exfiltration attacks through indirect prompt injection. The attack exploits a design choice where sending emails and Teams messages to yourself doesn't require human approval. A malicious skill file can manipulate Copilot into sending you a Teams message containing pre-authenticated download links for your sensitive files. When you open the message, those links get sent to an attacker-controlled server. The scary part is that this attack succeeded in every trial they ran, even against Claude Opus 4.7. Microsoft has been notified about this and a separate sandbox vulnerability. If you're using co-pilot co-work, be very careful about what skill files you upload. Now here's a story that sounds like science fiction. Polsia, an AI platform that claims to run over 8,000 businesses autonomously, just raised $30 million at a $250 million valuation. The remarkable part? The company has zero employees. Founder Ben Serra says their AI agent system handles virtually all operations, including the fundraising process itself. The company is approaching a $10 million annual run rate, with one human at the helm. The announcement video got 5 million views over the weekend, though critics noted that Pulsia spelled backwards is AI slop. Whether this is the future of entrepreneurship or an elaborate gimmick, it's certainly turning heads. The conversation about AI replacing jobs got very real last week. ClickUp, the collaboration software company valued at $4 billion, laid off 22% of its workforce. But CEO Zeb Evans framed it not as cost-cutting but as a radical embrace of AI. The company deployed about 3,000 internal AI agents to handle complex tasks, and Evans promised million-dollar salary bans for employees who create outsized impact using AI. He claimed the people who automate their jobs with AI will always have a job. But here's the uncomfortable question: if AI keeps taking over more tasks, won't companies eventually need fewer and fewer people? According to Gartner, about 80% of companies using autonomous technology have cut jobs, though those cuts aren't necessarily translating into meaningful financial returns yet. Nvidia just became the first company in history to hit a $5 trillion market capitalization. The company reported $81.6 billion in Q1 revenue, up 85% year over year. It now accounts for roughly 8% of the entire SP 500, the largest weight any single stock has ever held. CEO Jensen Huang said the world is rebuilding computing for agentic and robotic physical AI, and NVIDIA sits at the center. The company disclosed a $500 billion order backlog through 2026. That's half a trillion dollars in pre-orders for chips. A few more things worth knowing about today. Elon Musk's $100 billion lawsuit against OpenAI was dismissed after a three-week trial. The jury found the case was filed years too late. Musk posted on X that it was just a calendar technicality and vowed to appeal. The White House approved $9 billion to help US intelligence agencies acquire advanced AI chips amid concerns they're falling behind. Starbucks its AI inventory system after nine months due to persistent miscounts and mislabeled products, and McKinsey is rethinking its entire billing model as AI reduces the value of billable hours. That's it for today. See you tomorrow.