AI, Security, and Revenue: Anthropic Overtakes OpenAI

Today is one of those days when the AI world feels like a tech stage, a security conference, and a strategic poker game all at once. Between revenue milestones, new attack vectors, and fundamental debates about data and mathematics, one thing becomes very clear: AI is long past being just a model topic — it is now an infrastructure, business, and security topic.

And yes: if even Turing Award winners are arguing about how progress should even be measured, then you know things are getting serious.

Cisco has released security updates for several products, including BroadWorks, Crosswork Security, and Secure Workload. Especially concerning: according to heise, attackers can bypass login for Secure Workload in certain scenarios. That sounds dry, but in practice it is exactly the kind of vulnerability that gives admins sleepless nights.

Why does this matter? Because network and security products are often deeply embedded in core operations. If authentication can be bypassed there, it is not just “a bug,” but potentially access to sensitive infrastructure. For companies, that means: prioritize patches, check affected systems, and please do not wait for the next quiet week. Experience suggests it rarely arrives on its own.

This is also important for AI infrastructure, because AI stacks increasingly run on classic enterprise networking technology. Anyone ignoring security gaps there is building their shiny GPU world on shaky foundations. In other words: secure first, scale later.

💰 Anthropic overtakes OpenAI in revenue for the first time

Anthropic has overtaken OpenAI in revenue for the first time, as reported by The Decoder. That is not automatically a victory in the overall race, but it is a strong signal: monetization in the AI world is now almost as important as model quality. Whoever makes money faster has more room for compute, talent, and the next product wave.

This matters on several levels. First, it shows that the market is not only betting on the most recognizable brand, but apparently also on different product and sales models. Second, it makes clear how fierce competition in the LLM business is: benchmarks do not pay GPU bills. Third, it raises the question of whether OpenAI is currently going through a phase of strategic reorganization despite its enormous reach.

For you, this means: the AI market is not consolidating, it is accelerating. Revenue here is not just a number, but a proxy for product-market fit, enterprise adoption, and the ability to finance the next training run. The good old “move fast and break things” era now has a CFO.

🏭 AI lowers the barrier for attacks on industrial facilities

NSA, CISA, and the FBI are jointly warning that attackers are using AI to develop exploit scripts against industrial control systems. According to The Decoder, this affects Siemens S7 environments among others, and therefore critical sectors such as energy, water, and manufacturing.

The core of the warning: AI lowers the entry barrier. What used to require a lot of specialized knowledge, time, and experimentation can now be synthesized, adapted, and tested faster. That does not automatically make attacks more brilliant, but it does make them much more scalable. Especially in the ICS/OT environment, where systems often have long lifecycles and conservative update cycles, that is extremely dangerous.

For companies, this means: OT security no longer belongs in the “later” department. Network segmentation, patch management, monitoring, and incident response must also be realistically designed for industrial environments. Or in other words: the factory floor is now also an AI target. Unfortunately, without the luxury of simply “restarting.”

🌏 China has caught up in AI models

In the new edition of “KI-Radar,” The Decoder describes how Chinese models such as Kimi K3 and GLM-5.3 are now almost at U.S. level in benchmarks. There are also serious accusations that Chinese labs have systematically used Western models as “teachers.” Regardless of the question of guilt, the strategic message remains the same: the model lead is no longer a moat.

This is geopolitically relevant because AI leadership no longer depends only on research quality, but on chips, export controls, talent, access to data, and productization. If the pure model gap shrinks, competition shifts toward ecosystems, distribution, and infrastructure. That is exactly where Europe and other regions can still carve out niches — if they want to.

For the market, this means: we are moving from the question “Who has the best model?” increasingly to “Who can operate, integrate, and sell it best?” A small but important difference. Because in practice, the most elegant model does not always win — the one you can put into production tomorrow often does.

🧪 Sutton’s warning about synthetic data

AI pioneer Richard Sutton considers synthetic data a “big mistake” when it comes to scaling large language models, according to The Decoder. His argument: the world is infinitely complex, and simulations of it remain only tiny slices. Instead of feeding models ever more generated data, what is needed are agents that learn from real experience.

This is more than an academic disagreement. Synthetic data is seen by many as a pragmatic solution to data scarcity and exploding costs. Sutton’s criticism therefore hits a nerve: maybe LLMs are not simply “too small,” but their learning logic is ultimately limited. The alternative would be a shift toward continuously learning systems that are more strongly grounded in the real world.

This is relevant for AI teams because it affects the roadmap question: do you invest in data pipelines and post-training, or are you already building agentic systems with feedback loops? The answer is probably not either/or. But Sutton is a reminder that “more synthetic” does not automatically mean “more intelligent.”

🧭 Greg Brockman is expanding his influence at OpenAI

The Verge reports that Greg Brockman is continuing to expand his influence at OpenAI — precisely in a phase full of turbulence. The list is long: conflict with Elon Musk, an Apple trade-secrets lawsuit, internal departures, and the noise around a possible IPO. OpenAI currently feels less like a cozy research company and more like a tech giant in a permanent state of exception.

Why does this matter? Because leadership structure in AI companies is not just personnel policy, but product and strategy policy. When roles shift, the way decisions are made often changes too: faster, more centralized, or more conflict-ridden. For a company with enormous market influence, that can have direct consequences for roadmaps, partnerships, and capital market expectations.

The picture behind it is clear: OpenAI is continuing to professionalize, but under intense pressure. And the bigger the company gets, the more it needs not only good models, but also robust governance. Otherwise, “AI company” turns into “company with AI problems” faster than anyone would like.

📐 Terence Tao and a possible foundational crisis in mathematics

Mathematical genius Terence Tao sees the discipline facing a new foundational crisis, according to The Decoder — not necessarily logical, but cultural. The question is less whether a proof is true, and more how much human explainability, attribution, and appreciation remain in an AI-assisted proof world.

That is an interesting point, because with AI we often only look at output: did the model find the right answer? Tao reminds us that science is also a social practice. A proof that no human can follow may be formally correct — but what does that do to a discipline whose progress is based on comprehensibility and transmission?

For the AI debate, this is an important reality check. Not every increase in efficiency is automatically progress in the true sense. Or, to put it bluntly: just because the machine can do it does not mean the field will automatically be happier afterward.

🛠️ Tool tip of the day

If you want to keep a more structured eye on AI workloads, models, and infrastructure, it is worth taking a look at modern observability and experiment-tracking tools like #. Especially in the context of LLMs, GPU workloads, and agentic systems, you can quickly save time, money, and nerves. For teams that want to move from “it somehow works” to “it works reproducibly,” this is often the decisive step. #

Want to avoid missing any news? Subscribe to the newsletter