Tech & AI Unfiltered: The Smartest Twitter Threads You Missed Today

Anthropic’s new AI model found over ten thousand high-severity software vulnerabilities in one month. It’s a project called Glasswing, using an unreleased model named Claude Mythos, and it represents a permanent escalation in the cat-and-mouse game of cybersecurity. In our last episode, we talked about AI reinventing itself—and this week, we see the consequences, from the people building it, to the fundamental limits of the machines themselves. The game isn't just moving faster. The rules are being rewritten by the players themselves. And not all of them are on our team.

The most visible sign of that reinvention came with a major leadership shift. Andrej Karpathy—a co-founder of OpenAI and the former head of Tesla’s Autopilot—just joined Anthropic. He’s not going into a vague leadership role; he’s joining the pretraining team to work specifically on recursive self-improvement. This is the idea of AI improving AI. Karpathy has been publicly working on what he calls "autoresearch," and his move signals that Anthropic is now the place where the frontier of that work is being pushed. After the public drama of Musk versus Altman, one of the sharpest minds in the field has chosen his side.

He’s betting that the next great leap won't come from just making models bigger, but from making them smarter about how they build themselves. And while Anthropic attracts top talent, other platforms are demonstrating what AI can already do for pennies. An AI agent running on Coinbase’s Base network analyzed SpaceX’s one-point-five trillion dollar IPO filing in twelve minutes. It produced a full, institutional-style investment memo—analyzing valuation, Starlink revenues, and debt risks—for a total cost of one dollar and eighty-seven cents in crypto. For context, a single Bloomberg terminal, the gold standard for this kind of work, costs twenty-four thousand dollars a year.

The agent’s final recommendation on the IPO was a “Hold,” with specific conditions to buy. This isn't a toy. It’s a tool that just made a generation of junior investment bankers look very, very expensive. But as the capabilities expand, our understanding of the limitations is also getting sharper. A new paper called The Deterministic Horizon just dropped a bombshell on the entire field. Researchers from the University of Hong Kong have mathematically proven that transformer architectures—the foundation for everything from GPT-4o to Claude—have a hard ceiling on their accuracy.

For most models, that ceiling hits somewhere between nineteen and thirty-one layers. Beyond that point, performance doesn't just plateau. It decays. The paper states, quote, “No amount of additional training data, no choice of adapter rank, no sample size, and no loss function can push accuracy past this ceiling.” It’s not a data problem; it’s a structural, information-theoretic bottleneck. The "just add more layers and data" approach that defined the last five years of AI development just ran into a wall. And even for models operating below that ceiling, a new benchmark called Token Arena shows that performance is anything but guaranteed.

Researchers tested seventy-eight different endpoints for models like GPT-4o and Claude Sonnet. The results are shocking. The exact same model, deployed on different infrastructure, can have its accuracy vary by up to twelve-and-a-half points. The energy consumed per correct answer can be six times higher on one endpoint versus another. This means the leaderboards we all watch are telling, at best, half the story. The model you think you're using might be a finely tuned machine or a sputtering jalopy, depending entirely on the server it happens to be running on that day.

It complicates every claim of model superiority and puts the focus squarely on the unglamorous, hidden world of deployment infrastructure. Meanwhile, a pivot is happening in hardware. At the ISCAS conference, Huawei announced a new principle called the Tau Scaling Law. They’re proposing the industry stop chasing Moore’s Law—the obsession with shrinking transistors—and start focusing on reducing signal latency, the time it takes for data to travel. Using a new architecture called LogicFolding, they claim they can achieve the equivalent of one-point-four nanometer density by 2031, even without access to the world’s most advanced lithography machines.

This is a direct response to geopolitical restrictions. Instead of trying to break through the wall, Huawei is designing a way around it. It’s a strategic pivot that says if you won't let us build smaller, we’ll build smarter. Finally, while Anthropic’s new model was busy finding flaws in everyone else’s code, the company was dealing with a massive failure of its own. On March thirty-first, they accidentally leaked five hundred and twelve thousand lines of source code for Claude itself. This wasn't just any code. It was the logic for agent permissions and sandboxing.

The very systems designed to keep an AI in a safe, controlled environment were exposed for all to see. In a world where AI-powered offense is accelerating, this is the equivalent of handing over the blueprints to the bank vault. It’s the other side of the Glasswing story—and it’s where we need to look closer. Let’s go back to Project Glasswing. The headline number—ten thousand vulnerabilities in a month—is staggering. But the details are what matter. Anthropic gave their unreleased Claude Mythos model to fifty trusted partners. The model didn’t just scan for bugs.

It constructed working exploits. For example, it found a critical flaw in wolfSSL, a widely used security library, that allowed it to completely impersonate a secure HTTPS connection. This isn't a theoretical risk. This is a ready-to-use cyberweapon. The partners using Glasswing reported a tenfold increase in their bug detection rates. Anthropic’s own statement on this is chilling. They admit that, quote, “no company, including itself, has yet developed safeguards reliable enough to prevent malicious use” of these models. They’ve built the most powerful locksmithing tool in history, and they’re telling you, loud and clear, that they cannot guarantee it won’t be used to break into your house.

The process they describe is simple: "The model stays closed, the findings flow through coordinated disclosure, and the bug-fix bottleneck shifts to the human side of the loop." Read that last part again. The bug-fix bottleneck… shifts to the human side. AI can now find and weaponize exploits at a scale and speed that humans simply cannot match. The machine generates the threat. The human has to clean it up. This creates a profound asymmetry. The offense is automated. The defense is manual. And the defenders are already overwhelmed. Now, consider the other side of this.

At the exact same time Anthropic is demonstrating this god-like power to find flaws in complex systems, they leaked the source code for their own. Five hundred and twelve thousand lines of it. And what part of the code was it? The permission and sandboxing logic. The very brainstem of the AI that governs what it’s allowed to do, what files it can access, and what prompts it trusts. A cybersecurity expert, Tim Burke, put it perfectly. He said, quote, “Attackers got the entire blueprint for how an agentic AI validates permissions and handles credentials without having to reverse-engineer any of it.” Do you see the problem?

The two stories are the same story. On one hand, you have an AI that can accelerate the process of finding exploits to near-instantaneous speeds. On the other, you have a leak that gives attackers the exact instructions on how to bypass the safety systems of that very same class of AI. You are giving the automated burglar the blueprints to the alarm system. Microsoft, one of Anthropic’s largest partners, just canceled its internal licenses for Claude Code, citing cost pressures. Even at the highest levels, the cost-benefit analysis is getting complicated.

What is the price of a tool that can find ten thousand bugs for you, but might also create a single, unpatchable one in itself? The picture this week paints is one of immense, accelerating power paired with terrifying fragility. We have built engines of discovery that are operating at a level we can barely comprehend, let alone control. They are finding the cracks in our digital infrastructure faster than we can ever hope to patch them. And the companies building these engines are proving to be just as full of holes as the software they’re examining. We used to talk about the speed of innovation.

This week shows we need a new vocabulary. We need to talk about the speed of consequence. AI isn't just making things happen faster; it's collapsing the time between a breakthrough and a crisis, between a capability and a catastrophe. The insight from Glasswing—that the bottleneck is now human—is the single most important takeaway. The creation of exploits is now a scalable, automated process. The fixing of those exploits remains a messy, slow, human one. That gap between the speed of offense and the speed of defense is where the disasters of the next decade will live.

We are building tools that think and act on a timescale we can no longer keep up with. The question was never whether AI would be powerful. The question is what happens when that power is deployed into a world that is still, for now, run by humans who have to sleep, and eat, and manually patch the code. This week, we got a glimpse of the answer. The future is arriving faster than our ability to secure it.