OPENAI SOLVES 80-YEAR MATH MYSTERY, GITHUB GETS HACKED IN 18 MINUTES, TRUMP KILLS HIS OWN AI ORDER, AND ANTHROPIC CO-FOUNDER PREDICTS A NOBEL PRIZE
AN AI JUST SOLVED A MATH PROBLEM THAT STUMPED HUMANITY FOR 80 YEARS. THE MATHEMATICIANS ARE SAYING IT IS REAL THIS TIME.
SOURCE: TechCrunch | May 20, 2026
OpenAI announced that one of its general-purpose reasoning models independently disproved the Erdős unit-distance conjecture, a geometry problem first posed in 1946 that had stumped mathematicians for nearly eight decades. Unlike OpenAI’s embarrassing false claim from last year, this proof has been verified by multiple external mathematicians including Fields Medal-caliber researchers who have publicly endorsed the result.
Keywords: OpenAI math breakthrough, Erdos conjecture solved, AI reasoning model, AI mathematics 2026
Let me tell you about the last time OpenAI said they solved an Erdős problem. It was seven months ago, their VP posted on X with barely contained excitement that GPT-5 had cracked ten of Erdős’s most famous open problems, and the internet lost its mind for about 48 hours before several mathematicians calmly pointed out that GPT-5 had simply found existing solutions that were already in the literature. The VP deleted the post. Yann LeCun laughed. Demis Hassabis laughed. The rest of us sort of winced and moved on.
This time is different. And the remarkable thing is that the very same mathematicians who publicly humiliated OpenAI last year are now standing behind this result.
The problem in question is called the planar unit-distance problem, first posed by the legendary Hungarian mathematician Paul Erdős in 1946. The question is deceptively simple: if you scatter n points in a plane, what is the maximum number of pairs of those points that can be exactly distance 1 apart? For almost eighty years, the mathematical community believed the answer was essentially achieved by square grids. Nobody could prove it was optimal, but nobody could beat it either. It became one of those problems that a handful of researchers would poke at every decade, make marginal progress on, and then gently set back on the shelf.
An OpenAI reasoning model, not a system trained specifically on mathematics or targeted at this problem, just picked it up and broke it. The model discovered an entirely new family of geometric constructions using deep algebraic number theory, involving something called Golod-Shafarevich theory and infinite class field towers. It showed that square grids are not even close to optimal, and produced a polynomial improvement that external mathematicians have verified and extended. Princeton mathematician Will Sawin sharpened the bound to a specific delta of 0.014 within days of seeing the proof.
Thomas Bloom, who runs the Erdős Problems website and was the one who called OpenAI’s previous claim a dramatic misrepresentation, wrote in his statement this week: “AI is helping us to more fully explore the cathedral of mathematics we have built over the centuries.”
The implications here are genuinely strange to sit with. This is not a model that was told to solve this problem. It was a general reasoning model that connected ideas across fields in ways that human researchers had not explored. OpenAI says this means AI is now capable of sustaining long, difficult chains of reasoning that span multiple mathematical subfields. The same underlying capability has obvious implications for drug discovery, materials science, cryptography, and any discipline where the hard part is not doing the calculation but figuring out what to calculate in the first place.
Mathematicians have been cautiously celebrating. The rest of us are slowly realizing that the machines are now doing original mathematics. Whether that makes you feel excited or vaguely unsettled probably depends on what your PhD is in.
HACKERS RAIDED GITHUB IN 18 MINUTES THROUGH A POISONED VS CODE PLUGIN AND WALKED OUT WITH 3,800 INTERNAL REPOS FOR SALE
SOURCE: TechCrunch | May 20, 2026
GitHub confirmed it was breached by a hacking group called TeamPCP after attackers pushed a trojanized version of the Nx Console VS Code extension to the Visual Studio Marketplace, which remained live for just 18 minutes before being pulled. The malicious extension harvested credentials from developer machines including GitHub tokens, AWS keys, and Anthropic Claude Code configurations, ultimately enabling the exfiltration of approximately 3,800 internal repositories. The stolen code is now listed for sale on a cybercrime forum starting at $50,000.
Keywords: GitHub hack 2026, TeamPCP breach, VS Code extension malware, developer supply chain attack
Eighteen minutes. That is how long the poisoned VS Code extension was live on Microsoft’s Visual Studio Marketplace before someone noticed and pulled it. Eighteen minutes is roughly how long it takes to drink a cup of coffee. It is not very long at all, and yet it was apparently more than enough time for a hacking group to compromise an employee device at GitHub, steal credentials, and walk off with 3,800 internal code repositories containing the source code for GitHub Actions, Copilot, CodeQL, Codespaces, Dependabot, and various internal security tools.
GitHub is now telling its users that there is no evidence of impact to customer data stored outside of GitHub’s internal repositories. Which is technically reassuring until you think about what was inside those repositories. We are talking about the internal plumbing of the platform that houses essentially all of the world’s open source software. We are talking about the source code for the AI coding tool that millions of developers use every day. We are talking about internal security infrastructure whose exposure is the kind of thing that keeps security engineers awake at night wondering what comes next.
The group behind this, TeamPCP, is not some weekend hobbyist operation. They previously breached the European Commission, stealing over 90 gigabytes of cloud data. They hit Checkmarx. They hit Trivy, a vulnerability scanning tool, and used that foothold to reach downstream targets. This is a sophisticated outfit running a deliberate strategy: find a trusted tool in a developer’s workflow, poison it, and let the developer do the work of delivering the malware straight to the most valuable systems in their company.
The supply chain attack vector is the one that the security community has been screaming about for years, and it keeps working because the defense is genuinely hard. You cannot audit every VS Code extension you install. You cannot stop using developer tools. The attack surface is the nature of how software is built now, distributed and interconnected and dependent on thousands of upstream packages and plugins that you just have to trust.
What makes this particular attack notable beyond the target is the credential scope. The harvested data apparently included Anthropic Claude Code configurations, which means developer machines with Claude Code installed had their configurations and potentially their API keys sitting in the path of the exfiltration. That is a new dimension in AI-era security breaches. Your AI coding assistant is now part of your attack surface.
GitHub says they detected and contained it. The stolen repos are already listed for fifty grand on a cybercrime forum. The investigation is ongoing. If you are a developer and you have not rotated your GitHub tokens lately, this is a good week to do that.
TRUMP KILLS HIS OWN AI EXECUTIVE ORDER HOURS BEFORE THE SIGNING AFTER MUSK AND ZUCKERBERG CALL AND TELL HIM PLEASE DO NOT
SOURCE: Fortune | May 22, 2026
President Trump abruptly cancelled a scheduled signing ceremony for a major AI and cybersecurity executive order just hours before it was set to take place in the Oval Office, after Elon Musk, Mark Zuckerberg, and Trump’s own AI adviser David Sacks reportedly called him to lobby against the order. Trump cited concerns that regulation would dull America’s competitive edge on AI. The order was postponed indefinitely.
Keywords: Trump AI executive order cancelled, AI regulation 2026, David Sacks AI policy, Musk Zuckerberg Trump
Picture the scene. The Oval Office is set up for a signing ceremony. Tech and AI executives have been summoned to Washington. The cameras are ready. The pens are on the desk. And then, a few hours before showtime, the phone rings. Elon Musk is on the line. Mark Zuckerberg is on the line. David Sacks, Trump’s former AI czar who already quit once, is apparently back on the line too. And by the time those calls are finished, the executive order is dead.
This is the story of how America’s AI regulatory moment came, posed for a photograph, and then quietly left through the back door.
The order that Trump was about to sign was described as addressing AI safety and cybersecurity concerns, which sounds like exactly the kind of thing a government should be thinking about in the year 2026 when AI systems are solving 80-year-old math problems and hackers are poisoning developer tools to steal source code from GitHub. But the tech billionaires who have spent enormous amounts of money and political capital cultivating access to this administration were not interested in any framework that might slow their products down, and they made sure to say so before the ink was dry.
David Sacks, who served as Trump’s AI and crypto czar before stepping down in March, reportedly mounted a last-ditch lobbying effort by calling the president directly to express his concerns. Trump’s explanation for the cancellation was that he just did not like what he saw in the order’s text and that he worried it could dull America’s edge on AI. The phrase “America’s edge” is doing a lot of work in that sentence, because what it really means in practice is that any rule requiring AI companies to do anything that takes more than an afternoon will be framed as letting China win.
The irony here is considerable. The same week that GitHub got hacked through a poisoned developer plugin, and the same week that AI security vulnerabilities were in the news across multiple fronts, the administration pulled an executive order specifically designed to address AI and cybersecurity. The timing is either spectacularly bad luck or a very clear statement of priorities.
For what it is worth, Fortune notes that much of the MAGA base actually favors some form of AI regulation, particularly around issues like deepfakes and job displacement, putting the tech billionaire lobby at odds with the political coalition they helped build. That tension is going to get harder to paper over as AI systems become more visible in people’s daily lives and the consequences become more concrete.
The executive order will presumably be rewritten, softened, or simply forgotten depending on how the next news cycle goes. In the meantime, American AI policy remains whatever the last person to call the president told him it should be.
ANTHROPIC CO-FOUNDER TELLS OXFORD: AI WILL WIN A NOBEL PRIZE IN 12 MONTHS AND MIGHT ALSO KILL EVERYONE. HE SAID BOTH THINGS.
SOURCE: Time | May 22, 2026
Anthropic co-founder Jack Clark delivered the 2026 Cosmos Lecture at Oxford University, predicting that AI will help humans make a Nobel Prize-winning scientific discovery within 12 months, that bipedal robots will assist tradespeople within two years, and that AI-run companies generating millions in revenue will exist within 18 months. In the same breath, Clark maintained that plausible scenarios still exist in which AI poses a non-zero risk of killing everyone on the planet, and that this risk has not gone away.
Keywords: Jack Clark Anthropic Oxford, AI Nobel Prize prediction, AI existential risk 2026, AI scientific breakthrough
There is a very specific type of person who can stand at a podium at Oxford University and say “AI will help win a Nobel Prize within a year” and then, without changing his tone or his expression, say “and there is also a plausible scenario where it kills everyone on the planet.” Jack Clark is that type of person. He is one of Anthropic’s co-founders. He spent years at OpenAI before that. He has been in the room for essentially every major development in this industry for the better part of a decade. And his position, delivered at one of the world’s most prestigious lecture series, is essentially: this thing is going to be the greatest scientific tool in human history and we might also not survive it. Both. At the same time.
Let’s start with the optimistic part, because it is genuinely exciting. Clark’s prediction that AI will contribute to a Nobel Prize-winning discovery within 12 months is bold but not outlandish given what we already know. DeepMind’s AlphaFold team won a Nobel in chemistry in 2024 for protein structure prediction. OpenAI’s model just autonomously solved an 80-year-old mathematical conjecture this week. The same week Clark gave his lecture, there is corroborating evidence everywhere you look that AI systems are becoming capable of novel scientific contributions rather than just summarizing existing ones.
His other predictions are worth taking seriously too. Bipedal robots assisting tradespeople within two years, companies run entirely by AI generating millions in revenue within 18 months, and the genuinely alarming one: AI systems capable of designing their own successors by the end of 2028. That last prediction is the one that tends to make people’s eyes go a little wide, because it describes a feedback loop that is very difficult to reason about. If an AI can design a better AI, and that better AI can design an even better one, the question of what humans are for in that process becomes genuinely complicated.
And then there is the part about killing everyone. Clark was clear that he is not predicting this outcome. He is saying it is a plausible scenario with a non-zero probability, and that the existence of that probability means it has to be taken seriously even while building the product. This is Anthropic’s entire philosophy in a sentence. They believe they are building one of the most potentially dangerous technologies in human history and they are doing it anyway because they think it is better for safety-conscious people to be in the room than to cede the field to those who are less focused on safety.
Whether you find that logic compelling or deeply strange probably depends on how you feel about building things that might kill everyone. Most industries do not frame their work this way. Most people who build things do not start their pitch by saying “and there is a plausible scenario where this kills everyone.” Jack Clark does. He gave that speech at Oxford. He got a standing ovation.
Quantum Beat publishes daily. All stories sourced from verified publications. Views are editorial commentary.