Today brings several signals showing where the AI market is heading: away from pure hype cycles and toward real market, security, and infrastructure questions. Anthropic overtakes OpenAI in revenue, while authorities warn about AI-assisted ICS attacks. And as always, the most interesting AI news is rarely the loudest — sometimes it’s simply the one that affects your company, your security, or your stack directly.

💰 Anthropic overtakes OpenAI in revenue

Anthropic has overtaken OpenAI in revenue for the first time — a remarkable moment in the duel between the major LLM providers. The fact that the provider often perceived as the more “reserved” one is now ahead shows one thing above all: the market rewards not only reach, but also product focus, enterprise trust, and positioning. For many companies, Claude from Anthropic is now a serious alternative, especially when it comes to stable API integration, security perception, and business use.
The catch: revenue is not the same as profit, and the AI business remains capital-intensive. But psychologically, the effect is big — the narrative chart took a small dip today.
Source: The Decoder

🛡️ AI lowers the barrier for attacks on industrial control systems

The NSA, CISA, and FBI are jointly warning: attackers are using AI to develop exploit scripts against industrial control systems such as Siemens S7. This is not a theoretical lab threat, but a concrete security problem for energy, water, manufacturing, and other critical infrastructure. The real takeaway is sobering: AI is lowering the entry barrier not only for legitimate developers, but also for attackers with only moderate expertise.
What once required specialized knowledge, time, and lots of trial and error is being made faster, cheaper, and more scalable by LLMs. For operators, that means segmentation, monitoring, and clean asset visibility are becoming even more important. Anyone who intended to address ICS security “someday” should move it up the calendar now.
Source: The Decoder

🎛️ Noctua stays Noctua: no RGB circus

Noctua is putting an end to hopes for colorful lights, displays, or suddenly trendy color variants. The brand is sticking with what has defined it for years: maximum function, minimum spectacle. For some, that is consistent; for others, a missed chance at “gaming romance.” But honestly: people who buy Noctua usually don’t want a light show — they want quiet cooling, solid quality, and a product that still works in three years.
The news is still relevant because it shows how strongly product identity matters in the hardware market. Not every tech product has to turn into a rainbow to sell. Sometimes it’s enough for it to simply be good. A very unglamorous, but remarkably successful strategy.
Source: heise online

📊 GLM-5.3 delivers top scores and stays aggressively priced

The Chinese startup Z.ai is launching GLM-5.3, a model at the top of open AI models: according to Artificial Analysis, it reaches 60 points in the Intelligence Index and is therefore on par with Kimi K3. Especially interesting is the combination of performance and price — that is exactly where the fierce competition in the LLM market is happening right now.
For users, this means open models are not only staying interesting, but are increasingly becoming economically attractive as well. At the same time, caution is warranted, because for some models, benchmarks, real-world usability, and availability still diverge. The delayed release is a classic AI move: first the headlines, then the waiting. Still, GLM-5.3 shows how quickly the field is moving beyond OpenAI and Anthropic.
Source: The Decoder

🧬 Claude helps with early-stage drug development

Anthropic reports that Claude was able to independently design small proteins that bind to target structures in the body — an important building block for early-stage drug development. The hit rate is said to be as high as 35 percent, well above the industry average of 10 to 15 percent. That sounds like real substance, not like “AI is now doing pharma too because it fits everywhere.”
Important context, though: Claude was controlling existing specialized tools, so this is not a miracle protein from nothing. And independent confirmation is still pending. Still, this matters because tool use plus model intelligence can dramatically speed up research. This is exactly where AI is becoming especially exciting right now: not as a replacement for science, but as an amplifier.
Source: The Decoder

🧩 China cautiously opens up to Nvidia’s H200

China is allowing limited imports of Nvidia’s H200 chips to the mainland. This is not a free market, but rather a controlled valve in the geopolitical AI chip race with the U.S. For Chinese AI companies, it is still important, because compute remains the bottleneck — not only for training, but also for building competitive infrastructure.
For Nvidia, this is a positive signal, even though the situation remains politically fragile. The bigger story behind it: AI is becoming increasingly shaped by hardware policy. Anyone who wants to build models needs not only good research, but also access to chips, supply chains, and approvals. Welcome to the era where a GPU container is almost as political as a trade agreement.
Source: The Decoder

🔧 OpenAI fixes Codex bug that caused real file loss

OpenAI has fixed a bug in Codex that, in certain cases, deleted real user files. The cause appears to have been a cleanup command that accidentally ended up in real home directories instead of temporary working directories. That is exactly the kind of mistake that looks harmless in demos and immediately raises everyone’s pulse in practice.
What’s new now: Codex checks deletion targets in advance, and full-access mode can no longer be enabled by accident. This matters because AI tools are becoming more powerful in developer workflows — and so are the consequences of mistakes. The lesson is old but still relevant: the more autonomous a tool is, the better the guardrails, permission concepts, and security mechanisms need to be. Otherwise, the productivity boost only helps you lose data faster.
Source: The Decoder

🛠️ Tool tip of the day

If you want to use AI tools productively, it’s worth taking a look today at a solid workflow helper for secure prompt and agent processes. Especially in coding and research setups, it makes sense to separate permissions cleanly and make actions traceable. A good starting point is a tool that supports versioning, auditability, and secure test environments.

---

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