AI upends research, security, and chips in one day

Today makes it pretty clear where AI is really intervening right now: in science, in cybercrime, and in the semiconductor world. This is no longer some abstract vision of the future, but something happening now — with real consequences for research, security, and budgets. And yes, regulation and training are also lagging behind, usually at a somewhat slower pace than the models.

🧪 Two teams, one crypto problem, one model

Two independent research teams solved the same open problem in quantum cryptography — almost simultaneously and both with the help of OpenAI’s GPT-5.6 Sol Ultra. The spectacular part is not just the solution itself, but the fact that both groups apparently arrived at the same AI workflow and published just three hours apart. That is a pretty hard reality check for the idea that scientific discovery is always a linear, purely human flash of inspiration.

Why this matters: If AI models like GPT-5.6 are not only assisting but actively sparking new proofs or derivations, the logic of research changes. Then it is no longer just about “AI as a tool,” but about what originality, independence, and priority even mean in research anymore. For science, that is exciting — and for anyone who believes in “the one genius stroke,” slightly unsettling. Source: The Decoder

🕵️ AI is becoming cybercrime infrastructure

Interpol reports that AI is now involved in 55 percent of reported cybercrimes in Africa. Financial damage rose from 192 to 484 million US dollars, and around 600,000 cases of digital extortion involving deepfakes were also recorded. This is not just a statistic, but a sign that AI is long since supporting the entire chain of cybercrime: from deception and phishing to extortion and fraud.

For companies and public authorities, this matters above all because classic security measures against social engineering are reaching their limits. If a language model writes convincing phishing emails or creates deepfakes for CEO fraud, “look again carefully” is no longer enough as a security strategy. At the same time, the report shows that regulation and awareness campaigns fall short without technical defenses. In short: attackers are scaling with AI — defenders have to do the same. Source: The Decoder

🧠 What is really transferable when compressing LLMs

A new arXiv paper examines what really transfers between Transformer models of different sizes — and what does not. The researchers characterize a 1.4B-to-410M conversion in the Pythia family and reach an important finding: representations can be aligned strongly, but parameters only weakly. In other words: the model still “knows” some things in a similar form, but blindly projecting the weights destroys functionality more than it elegantly shrinks it.

This is relevant for LLM infrastructure and inference optimization, because many teams want to derive smaller, cheaper models from large ones. If that does not work cleanly, distillation, projection, or other compression methods quickly become more expensive or riskier than expected. For anyone working with on-device AI, edge deployments, or cheaper inference, this is an important reality check. The good news: there is apparently structure you can use. The bad news: it is not magic. Source: arXiv

🏛️ NIS2: registration is underway, but many are still missing

Heise reports on the mystery surrounding missing NIS2 registrations: the grace period for operators of critical facilities expired on 2026-07-31, but the number of registered companies remains below expectations. That sounds like bureaucratic aftertaste at first, but in practice it is quite relevant. After all, NIS2 is one of the central European cyber security regulations and significantly raises the bar for organizational and technical protective measures.

For companies, this means: if you operate critical infrastructure or are tied into corresponding supply chains, you should not assume that “we discussed the topic internally once” replaces any form of compliance. The report also shows how difficult implementing regulatory requirements remains in reality — especially when responsibilities are unclear or registration has simply been delayed. Less glamorous than GPT-5.6, but probably with more immediate consequences. Source: heise online

⚙️ AMD benefits from the AI wave

AMD has delivered strong numbers in its latest quarterly report: revenue in the data center business came in at 6.7 billion US dollars and more than doubled year over year. The driver is the unabated demand for AI capacity, while the gaming business moves into the background. That is a pretty clear signal of where hardware demand is shifting: data center chips are the real headline act right now, not the prettier graphics card for the desk.

For the market, this means: the AI economy remains a hardware game. Whoever delivers compute, memory bandwidth, and efficient infrastructure wins. For users and companies, it means: more competition in AI chips is good, but supply of compute remains expensive and strategic. AMD’s numbers also show that Nvidia is not the only beneficiary of the boom — even if the market leader probably has no intention of giving up the throne voluntarily. Source: The Verge

🛠️ Tool tip of the day

If you work on cybersecurity in production environments, training is right now not a “nice to have,” but a must. The Hanover University course on IT security in production facilities teaches the fundamentals of IT and OT security as well as standards and norms for secure plant design. For teams with interfaces between production, operations technology, and IT in particular, a format like this can help close the typical blind spots.

What is practical about it: you do not just get theory, but a better understanding of how to think about security in industrial environments at all — beyond “we’ll install one more tool and hope for the best.” #

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