AI Between Space, Politics, and Security Vulnerabilities
AI news of the day: Google tests data centers in space, US politics wants to regulate superintelligence, and new security risks are adding pressure.
Inhaltsverzeichnis
Today brings an unusually dense mix of hype, risk, and regulation. From AI data centers in outer space to new US legislative plans and security vulnerabilities in models that are supposed to help, one thing is becoming clear once again: the AI industry is not only building products, but also its own guardrails — often at the same time.
And as so often, the most interesting questions are not just what is technically possible, but who pays the bill, who is liable, and who can still make sense of it all in the end. So: take a deep breath, then dive into the news.
🌍 Google tests AI data centers in space
With Project Suncatcher, Google is thinking big — literally. The company wants to send its first AI satellites into space in order to build computing capacity outside Earth over the long term. On October 1, an initial satellite roughly the size of a refrigerator is even set to launch aboard a SpaceX Falcon 9. The idea sounds like science fiction, but it is a real infrastructure test: in space there would be more solar energy and, in theory, better cooling than on Earth.
The catch, however, is enormous. For a single 1-gigawatt data center, estimates suggest around 10,000 satellites would be needed. That is not an “MVP” — that is more like a space program with spreadsheets. Still, Suncatcher is relevant because it shows how urgently the industry is looking for new ways to meet AI’s exploding energy demand. If you want to know where AI infrastructure is headed: not just to the cloud, but apparently also toward orbit.
Source: The Decoder
🏛️ US draft law aims to ban superintelligence
In the US, AI regulation is now becoming even more direct: Senator Bernie Sanders and Representative Greg Casar have introduced a draft that would permanently ban the development and use of artificial superintelligence. At the same time, a new federal AI agency is being discussed. This is politically notable because the debate is no longer just about transparency or liability, but about a possible ban on an entire technological field.
In practice, the draft is mainly a signal: the debate around AI safety, control, and concentration of power has finally entered the mainstream. Even if a ban in this form is unlikely to pass, it shifts the frame. Companies will have to prepare more seriously for regulatory limits — and not just in Europe. For startups relying on large models, agents, or automation, this means compliance is no longer a side issue, but a product factor.
Source: The Decoder
🔐 Meta Muse is said to expose the entire filesystem
Security researchers are reporting a rather embarrassing but instructive finding: Meta’s AI Muse is said to be coaxed with little prompting into revealing the entire filesystem — including Ubuntu system files, app templates, and internal documentation. If confirmed, this is more than a minor bug. Then we are talking about a massive problem in model safety, isolation, and access control.
Why does this matter? Because many AI products today no longer just “chat,” but work directly with tools, filesystems, and internal processes. That is exactly where the most dangerous failures happen: a model is supposed to help efficiently, but one wrong permission and assistance turns into a data leak. For companies, it is a reminder that prompt design alone is not a security strategy. And for users, the somewhat unromantic truth: an LLM is not a vault just because it answers eloquently.
Source: The Verge
📈 Experts continue to underestimate AI progress
A new analysis by the Forecasting Research Institute shows that even top experts have systematically underestimated AI progress in recent years. There are plenty of examples: gold-level performance in the Mathematics Olympiad was reached about five years earlier than expected, and Anthropic’s revenue was far above the highest estimates. This is interesting because it reveals a recurring pattern: AI develops especially quickly where scalable training data, compute, and clear benchmarks come together.
At the same time, the picture remains mixed. In autonomous systems or self-driving, the results are much less spectacular. So the right takeaway is not: “Everything is moving faster than expected.” Rather: we often overestimate the linear continuation of old trends and underestimate how abruptly some areas can evolve. For investors, product teams, and policymakers, that is an uncomfortable but important insight. Forecasting AI is necessary — just often with built-in humility.
Source: The Decoder
🎙️ ElevenLabs speaks openly about bots and margins
ElevenLabs is long past being just a cool voice startup and is now a serious player in AI infrastructure for speech. In an interview with TechCrunch, the CEO explains, among other things, that companies should ideally inform customers when they are speaking with a bot — at least as long as that is not yet self-evident. That is more than PR: it is a small ethics-and-product question with a big impact.
Why is this relevant? Because voice AI is becoming more and more common in customer service, sales, and assistive systems. Exactly there, transparency determines trust. If the voice sounds human but a model is actually speaking, the line between convenience and deception becomes thin. At the same time, the example also shows how the business model is shifting in the market: it is not only quality that matters, but also margins, IPO timing, and how quickly the market accepts bot communication. Anyone building LLM products should take a close look here.
Source: TechCrunch
🧠 Gemini 4 is probably coming sooner than expected
At Google DeepMind, the focus is clearly shifting toward product. Head Koray Kavukcuoglu says Gemini 4 is coming “much earlier” than the end of the year; the model is already in post-training and is running internally in the coding tool Antigravity. At the same time, the old AGI debate seems to matter less to him — what matters more are “trustworthy agents.” That is notable because it describes the tone across the entire Google ecosystem: less philosophical end-of-the-world debate, more market-ready features.
For the industry, this is a signal that the race is not only about model size, but about concrete workflows: coding, agents, product integration. If DeepMind really becomes Google’s product department, that also means quality must translate into real usage faster. And that is exactly where research and delivery diverge. Or put differently: less “When will AGI arrive?”, more “How do we get the feature into the app by Tuesday?”
Source: The Decoder
🛠️ Tool tip of the day: take vibe-coding seriously
If you want to keep an eye on developments around vibe-coding, rapid prototyping, and productive LLM workflows, it is worth using a tool that lowers the barrier from prompt to app. Especially in the world of no-code, agents, and startup prototyping, such tools are often the difference between “interesting idea” and “already running internally.” #
Don’t want to miss any news? Subscribe to the newsletter