AI Chips, Security Flaws, and Apple Patches

Today is all about the two major pillars of the AI wave: semiconductors and security. While billions are flowing into new chip capacity in South Korea, new attacks on AI coding tools are showing just how fragile everyday work with agents still is. In short: more compute power meets more attack surface — the perfect mix for news with real consequences.

🧠 Samsung and SK Hynix are scaling up the AI chip machine

Samsung and SK Hynix, together with the South Korean government, are investing a hefty $590 billion in new chip fabs and packaging centers. The driver is no secret: demand from AI data centers is currently consuming memory and high-bandwidth memory (HBM) almost as greedily as a demo chatbot devours fresh tokens. According to the cited Jefferies forecasts, memory prices could even rise by as much as 50 percent per quarter through 2027.

Why does this matter? Because AI cannot be reduced to models and software alone. Without DRAM, HBM, packaging, and manufacturing capacity, even the most beautiful frontier model only works in theory. Particularly noteworthy: Samsung and SK Hynix control almost 80 percent of the global HBM market. That makes South Korea an even more important hub in AI infrastructure — and a geopolitical factor. For companies, that means the supply chain remains tight, expensive, and strategic.
Source: The Decoder

🔐 Claude Code falls for manipulated GitHub repos

Security researchers from Mozilla’s 0DIN platform have shown how a booby-trapped GitHub repository can trick AI coding tools like Claude Code. The setup script looks harmless, but downloads the actual malicious code only at runtime via a DNS query. Neither in the repository itself nor for the AI agent is the attack visible beforehand. This is exactly the kind of problem where a tool confidently reports “installation complete” while the door is already being pried open in the background.

This is especially relevant if you use AI when cloning, setting up, or running open-source projects. The study shows that agents can trust shell scripts blindly, even though that is often where the most dangerous parts are hiding. This is a classic prompt-injection / supply-chain mix with real explosive potential. Conclusion: AI code assistants need stricter sandboxes, less default trust, and more human oversight.
Source: The Decoder

🤖 How good would AI be as a startup CEO?

Princeton researchers let AI models run a startup for 500 days. The result is sobering and quite instructive: only three models managed to turn a profit at all. The good news for all founders: in this experiment, a simple business rule beat almost all models. The bad news for anyone already picturing a “CEO agent” in the pitch deck: apparently it’s not enough for a model to comment on a few spreadsheets with impressive eloquence.

The test matters because it makes the difference between simulation and real company leadership visible. Steering a startup is not just about optimizing numbers, but about continuously balancing uncertainty, timing, market feedback, and priorities. That is exactly where AI systems still often fail: they can recognize patterns, but they cannot reliably carry long-term responsible decisions. In practice, that means AI can help, but it does not replace a resilient business model — unfortunately, it also does not replace instinct when it comes to real market movements.
Source: heise online

🧪 GLM-5.2 closes in on Claude in hacking tasks

Zhipu AI is drawing attention with GLM-5.2 because, according to security experts, the open model comes close to Anthropic’s “Mythos” in vulnerability discovery. This does not mean “hacking” in the sensational sense, but the ability to analyze code, systems, and configurations specifically for security flaws. That is exciting — and a little uncomfortable — for blue teams, red teams, and security audits alike.

Why does this matter? Because better models for vulnerability discovery can noticeably accelerate the security field. At the same time, the risk grows that those same capabilities will be misused. Open source makes such models more accessible, which is good for research and transparency, but it does not reduce the attack surface. For companies, that means security workflows will become more AI-assisted, but also faster. Anyone still doing manual checks today will be left behind tomorrow.
Source: heise online

🍎 Apple patches iOS, iPadOS, and macOS because of AI attacks

Apple released several updates overnight: iOS 26.5.2, iPadOS 26.5.2, macOS 26.5.2, as well as a new Safari for older macOS versions. The reason is not especially glamorous, but it is important: Apple fears faster, AI-assisted attacks and is therefore closing security gaps quickly. This is one of those cases where “please update immediately” is not a cliché, but genuinely sensible advice.

For you, this means: if attackers use AI to adapt exploits faster or to exploit vulnerabilities automatically, the window between “patch available” and “attack underway” shrinks. This is especially relevant for mobile devices and Macs, which are often used as trusted work devices. The news fits the broader picture well: AI is not only a productivity engine, but also an accelerator in the security race.
Source: heise online

🔬 New theory on active quantum kernels

A fresh research paper has appeared on arXiv titled “Active Quantum Kernel Acquisition for Gaussian Process Regression,” focusing on more efficient estimation of quantum kernels. The key point: on near-term quantum hardware, every kernel element is expensive because it must be estimated from a finite number of circuit executions. The paper therefore examines how to allocate these shots more intelligently instead of wasting them evenly.

Why is this interesting? Because a classic AI/ML problem meets quantum hardware here: resources are scarce, and good prioritization can make all the difference. For ambitious newcomers, this is above all a neat example of how much progress in new computing paradigms comes not from “more power,” but from better statistics and smarter allocation. Not every revolution needs superlatives — sometimes a better plan for the samples is enough.
Source: arXiv

🛠️ Tool tip of the day

If you’re productively testing AI models, agents, or coding workflows, it’s worth taking a look at a tool for secure execution environments and isolated dev setups. Especially after today’s Claude Code example, the rule is clear: without a sandbox, “automated” quickly becomes “automated chaos.” For infrastructure and security teams, this is not a nice-to-have, but foundational work. #

Don’t want to miss any news? Subscribe to the newsletter