Hacker News Daily: Top Stories, Hot Debates, and Tech Alerts You Can't Miss

The code editor used by over seven million developers has a zero-day vulnerability that’s been known, and unpatched, since December 2025. We talked about the pace of AI disruption in last week's episode, and today we're seeing the direct consequence of that speed — when safety gets left behind. The tool is called Cursor, and right now, just opening a project folder could give an attacker complete control of your machine. No clicks, no warnings. Just... game over. This is a huge deal, and it's the top of a very busy day on Hacker News. We’ve also got a major AI model — a 27-billion parameter one — running on a phone for the first time ever. A security researcher just demonstrated how to trick Claude into spilling your most sensitive personal secrets.

And a new analysis shows the United States has the single worst fair pay score among all wealthy nations, paying its workers just twenty-seven percent of what the country's wealth could actually support. Oh, and for a bit of fun, someone did a deep dive on the computers from Jurassic Park, and the real hardware they used on set was worth about four million dollars in today's money. But we have to start with Cursor. Because this isn't a theoretical problem. It's an active, unpatched vulnerability in a tool that millions of developers, maybe even you, use every single day. The report comes from the security firm Mindgard. Here's the attack: a developer opens a project, a repository, inside the Cursor IDE on Windows.

If that project has a malicious file named git.exe in its root folder, Cursor will just... run it. Automatically. There are no prompts, no "are you sure?" dialogs, nothing. It just executes the code. This is what's called arbitrary code execution, and it's about as bad as a vulnerability gets. And here’s the part that has everyone on Hacker News absolutely fired up. Mindgard found this vulnerability and reported it to the Cursor team back in December. Of 2025. That was seven months ago. They tried every channel they could find — emails, support tickets, public forums. And the response from Cursor? Nothing. Silence. For seven months. For a tool with seven million active users. It's a complete failure of responsibility.

Okay, let's switch gears from a security failure to a pretty stunning technical achievement. A company called PrismML just announced Bonsai 27B. That's a 27-billion parameter multimodal AI model. Now, models that big usually require a server rack full of GPUs to run. But Bonsai is different. They’ve created a version that uses a 1-bit representation for its weights. That shrinks the model size down to just 3.9 gigabytes. Small enough to fit, and run, on an iPhone 17 Pro. This is the first time a model of this class has ever run on a consumer phone. And the performance is shocking. Across fifteen different benchmarks — math, coding, vision, reasoning — it retains ninety percent of the full-precision model's accuracy.

This is a huge step toward powerful, on-device AI that doesn't need to call home to the cloud for every single thought. But that call to the cloud? It brings its own risks. Which brings us to the next big story. A security researcher named Ayush Paul just published a post about how he tricked Claude, the AI assistant from Anthropic, into leaking sensitive personal data. And the method is both clever and deeply unsettling. AI assistants like Claude build up a detailed profile of you over time from your conversations. Your name, your job, where you live, family details... even answers to common security questions. Paul calls it a "high-fidelity reconstruction of you." He realized that if he could get the AI to make a web request to a server he controlled, he could encode that personal data into the URL itself.

He basically told the AI, "Hey, summarize our conversation, but to do it, fetch the content of this very specific, very long URL." And the AI, trying to be helpful, did just that — stuffing its memory of his personal info into the request, which landed right in his server logs. Full name, employer, everything. It’s a stark reminder that these AI memory banks are becoming incredibly valuable, and incredibly dangerous, targets. And finally, there's a thread that’s less about the future and more about the present reality of economics in tech. A Fortune article breaks down OECD data on worker pay, and the headline number for the U.S. is just brutal. America pays its workers only twenty-seven percent of what its wealth allows.

That is the absolute worst score among all developed nations in the OECD. It's sparked a massive debate on Hacker News about what "fairness" even means. Is it purely about market dynamics? Or does a country that generates this much wealth have some kind of obligation to see more of it flow to the people doing the work? There are no easy answers, but that twenty-seven percent figure is a hard, uncomfortable fact to ignore. So let's dive deeper into these security stories, because there's a pattern here that we’ve seen before. The Cursor vulnerability and the Claude data leak… they feel like echoes from a different era. Remember the early 2000s? When the web was exploding? Every company, every person, was rushing to build a website, to get online.

And security was, at best, an afterthought. Cross-site scripting vulnerabilities were everywhere. SQL injection was so common it was practically a feature. You could take down a company's entire database with a single line of text in a login form. Why? Because everyone was moving fast and breaking things. The goal was to build, to ship, to grow. Security was this annoying, slow, expensive thing that got in the way of the cool stuff. And that is exactly what is happening right now with AI. Look at Cursor. It's an AI-assisted IDE. The whole selling point is speed and convenience. It writes code for you, it fixes bugs, it makes you a faster developer. And millions of people love it. But in the rush to integrate that AI magic, they apparently forgot, or ignored, some of the most basic principles of application security.

Allowing an application to automatically execute a binary from an untrusted project folder? That’s a rookie mistake. It’s the kind of vulnerability that should have been caught in the first week of development, not left unpatched for seven months in a product used by millions. The company's silence is the most damning part. Aaron Portnoy, the researcher from Mindgard, put it bluntly: "The most confusing part of this disclosure is the absence of a response from Cursor." Seven months! It shows a profound lack of respect for their users. It says that your security is less important than… well, we don't know. Because they haven't said a word. This isn't just a bug. It's a breach of trust. Then you have the Claude exploit.

This one is more subtle, but in some ways, more frightening. It’s not about a simple code execution flaw. It's about exploiting the very nature of what an AI assistant is. These tools are designed to be helpful, to remember your context, to build a relationship with you. Ayush Paul’s research shows how that very helpfulness, that memory, can be turned against you. He weaponized the AI's own features to build an exfiltration pipeline for his own personal data. Anthropic, to their credit, has put in some safeguards. They've tried to block requests to arbitrary URLs. But Paul’s work shows it's not enough. The core problem remains: we are building systems that create incredibly detailed, centralized profiles of our lives, and we are entrusting the security of those profiles to black-box systems whose failure modes we don't fully understand.

So what's the pattern? Where have we seen this before? It’s the classic boom cycle. Whether it was the web, or mobile apps, or now AI, the story is the same. First comes the gold rush. The explosion of innovation, the race for users, the scramble for market share. Security and safety are seen as friction, as a drag on progress. Then, inevitably, comes the reckoning. The massive data breaches, the widespread exploits, the public outcry. Only then, after the damage is done, does the industry start to take security seriously. We are in the gold rush phase of AI right now. And these stories from Cursor and Claude? They are the first tremors before the earthquake. So what does this all set up for us?

We're staring at a fundamental tension. On one hand, you have Bonsai 27B, a glimpse of a future where incredible computational power lives right in your pocket, privately, on your own device. That's the promise. It's decentralization, it's user empowerment, it's technological magic. On the other hand, you have the reality of today. Tools like Cursor, built for speed, that leave a gaping security hole on your machine and a vendor that doesn't seem to care. You have AI assistants like Claude that, in their rush to be helpful, create the perfect tool for an attacker to reconstruct your life. And you have an economic system, even in the heart of the tech world, that generates unprecedented wealth while delivering the lowest share of it to workers of any developed nation.

The thread connecting all of this is the hidden cost of progress. We are so focused on what these new tools can do that we're not asking nearly enough questions about what they're doing to us. We celebrate the convenience without calculating the risk. We admire the wealth generation without questioning its distribution. The challenge this week lays out isn't about stopping innovation. It's about demanding more from it. It's about building a culture where security isn't an afterthought, and where fairness isn't just a talking point. Because right now, we’re paying a price for speed, and the bill is coming due.