Anthropic admits that Claude was misused for missile guidance software, warns shortly afterwards about AI models that copy themselves onto cheap devices, and the EU postpones the high-risk obligations of the AI Act to the end of 2027. Why AI regulation can no longer wait.
Today I want to talk again about the latest developments in AI, and specifically about the uncomfortable ones. It is about how AI keeps slipping further out of control, how the big corporations apparently no longer have it under control themselves, and why these are risks that affect all of us as users. Three reports from the space of a few weeks show the pattern.
Recorded in Frankfurt am Main, October 2026. All sources with timestamps are in the video description.
Claude in a missile program: a corporation admits misuse
In its September threat intelligence report, Anthropic describes how a group in northern Yemen used Claude Code for missile guidance and control software. The attribution to the Houthis comes from media reports, Anthropic itself speaks of a cell in the north of Yemen. According to Anthropic, ballistic data was aggregated, processed via flight simulators and real-time data, and used for flight. By the time Anthropic recognized how the software was being used, the attackers had already been able to apply what they had learned and make adjustments.
For context: according to Anthropic’s account, there is no evidence that an operational weapon was completed, a test flight apparently failed, and the accounts were suspended. That does not change the core issue. In a rapidly changing environment, an AI can calculate better and react faster than any human. Exactly this capability was turned against people here, and the manufacturer only noticed after the fact.
Who controls the AI corporations?
For me, that is the real question. Nobody controls what Anthropic, OpenAI, Meta with Llama or whoever else out there is developing. There is no authority that says: wait, this cannot go in, we are slowing down here. Today it is control data for missiles, tomorrow it could be instructions for things I do not even want to spell out here. And this is only the beginning.
AI that copies itself onto household devices
On the same day I recorded the video, Anthropic’s red team published a warning: smaller AI models could in the near future be able to replicate themselves autonomously onto cheap IoT devices, without a data center. This was tested on inexpensive TV boxes. Anthropic itself writes that there is no complete replicator yet and that the tests ran in simplified environments. But the direction is clear.
Think about what that means. Almost every device today contains a small computer. If an AI can copy itself there, how are you ever going to switch it off again? And if it sneaks into a safety-critical device and an error happens there: who is to blame, who bears the responsibility? That sounds like a bad movie, but even the manufacturer is warning about it.
Why AI regulation is so slow
Regulation is moving far too slowly, and that is no coincidence. No large corporation wants AI to be regulated. Regulation slows progress, less progress means less competitiveness, and whoever is not competitive will not make money tomorrow. According to netzpolitik.org, Big Tech puts around 151 million euros a year into EU lobbying. Profitability is placed above control, above safety and, if in doubt, above human lives. That is sad, but it explains why the EU has just postponed the high-risk obligations of the AI Act to 2 December 2027 with the Digital Omnibus.
What would actually be simple: seal, registration, control
Register AI. Every model that is placed on the market belongs in a register, just as the AI Act already provides for high-risk systems.
Issue a safety seal. A quality and safety seal that clearly says: this model has been tested and you can use it, this one not.
Regulate deployment. Clearly define which AI may be used when and for what, and enforce it too.
If that does not happen, we will arrive faster than expected at a point where AI is uncontrollable and we can no longer go back to register, control or regulate. Then our data will be in every AI, our digital footprints will be everywhere, and the question of what humans are still needed for will at some point no longer be asked rhetorically. How companies build this control internally is something I described in my post on agentic AI and AI governance (German).
EU AI Act and Germany: slower, but with a head on our shoulders
Fortunately, here in Europe we have a foundation with the EU AI Act that exists nowhere else in the world: introduce AI only after it has been regulated and controlled. Yes, in Germany we are slower in development than the competition. But in the end, Germany is the country that delivers quality and that everyone buys from, because here things are built with a head, with sense and with regulation. That is exactly how truly safe systems come about. What the AI Act specifically requires of banks and insurers is in my post on Agentic AI governance under BaFin and the EU AI Act (German).
I stand by it: we have to start controlling AI. What do you think, is the AI Act enough, or do we need a seal for every model? Let me know on LinkedIn or under the video on my YouTube channel.
Frequently asked questions
Did Anthropic confirm that Claude guided missiles?
In the report, Anthropic describes that a cell in northern Yemen used Claude Code for guidance and control software, but cites no evidence of an operational weapon. The attribution to the Houthis comes from media reports. The accounts were suspended.
What does the postponement of the high-risk obligations in the AI Act mean?
With Regulation (EU) 2026/1744, the Digital Omnibus on AI, the obligations for high-risk systems only apply from 2 December 2027, and for systems under Annex I from August 2028. Companies gain time, but they do not lose the obligation.
What can a company do today?
Keep its own AI register, deploy only tested models, define clear rules for use and let humans intervene at the decisive points. That is AI governance on a small scale, and it works regardless of when the legislator is finished.
Anthropic admits that Claude was misused for missile guidance software, warns shortly afterwards about AI models that copy themselves onto cheap devices, and the EU postpones the high-risk obligations of the AI Act to the end of 2027. Why AI regulation can no longer wait.
Today I want to talk again about the latest developments in AI, and specifically about the uncomfortable ones. It is about how AI keeps slipping further out of control, how the big corporations apparently no longer have it under control themselves, and why these are risks that affect all of us as users. Three reports from the space of a few weeks show the pattern.
Claude in a missile program: a corporation admits misuse
In its September threat intelligence report, Anthropic describes how a group in northern Yemen used Claude Code for missile guidance and control software. The attribution to the Houthis comes from media reports, Anthropic itself speaks of a cell in the north of Yemen. According to Anthropic, ballistic data was aggregated, processed via flight simulators and real-time data, and used for flight. By the time Anthropic recognized how the software was being used, the attackers had already been able to apply what they had learned and make adjustments.
For context: according to Anthropic’s account, there is no evidence that an operational weapon was completed, a test flight apparently failed, and the accounts were suspended. That does not change the core issue. In a rapidly changing environment, an AI can calculate better and react faster than any human. Exactly this capability was turned against people here, and the manufacturer only noticed after the fact.
Who controls the AI corporations?
For me, that is the real question. Nobody controls what Anthropic, OpenAI, Meta with Llama or whoever else out there is developing. There is no authority that says: wait, this cannot go in, we are slowing down here. Today it is control data for missiles, tomorrow it could be instructions for things I do not even want to spell out here. And this is only the beginning.
AI that copies itself onto household devices
On the same day I recorded the video, Anthropic’s red team published a warning: smaller AI models could in the near future be able to replicate themselves autonomously onto cheap IoT devices, without a data center. This was tested on inexpensive TV boxes. Anthropic itself writes that there is no complete replicator yet and that the tests ran in simplified environments. But the direction is clear.
Think about what that means. Almost every device today contains a small computer. If an AI can copy itself there, how are you ever going to switch it off again? And if it sneaks into a safety-critical device and an error happens there: who is to blame, who bears the responsibility? That sounds like a bad movie, but even the manufacturer is warning about it.
Why AI regulation is so slow
Regulation is moving far too slowly, and that is no coincidence. No large corporation wants AI to be regulated. Regulation slows progress, less progress means less competitiveness, and whoever is not competitive will not make money tomorrow. According to netzpolitik.org, Big Tech puts around 151 million euros a year into EU lobbying. Profitability is placed above control, above safety and, if in doubt, above human lives. That is sad, but it explains why the EU has just postponed the high-risk obligations of the AI Act to 2 December 2027 with the Digital Omnibus.
What would actually be simple: seal, registration, control
If that does not happen, we will arrive faster than expected at a point where AI is uncontrollable and we can no longer go back to register, control or regulate. Then our data will be in every AI, our digital footprints will be everywhere, and the question of what humans are still needed for will at some point no longer be asked rhetorically. How companies build this control internally is something I described in my post on agentic AI and AI governance (German).
EU AI Act and Germany: slower, but with a head on our shoulders
Fortunately, here in Europe we have a foundation with the EU AI Act that exists nowhere else in the world: introduce AI only after it has been regulated and controlled. Yes, in Germany we are slower in development than the competition. But in the end, Germany is the country that delivers quality and that everyone buys from, because here things are built with a head, with sense and with regulation. That is exactly how truly safe systems come about. What the AI Act specifically requires of banks and insurers is in my post on Agentic AI governance under BaFin and the EU AI Act (German).
I stand by it: we have to start controlling AI. What do you think, is the AI Act enough, or do we need a seal for every model? Let me know on LinkedIn or under the video on my YouTube channel.
Frequently asked questions
Did Anthropic confirm that Claude guided missiles?
In the report, Anthropic describes that a cell in northern Yemen used Claude Code for guidance and control software, but cites no evidence of an operational weapon. The attribution to the Houthis comes from media reports. The accounts were suspended.
What does the postponement of the high-risk obligations in the AI Act mean?
With Regulation (EU) 2026/1744, the Digital Omnibus on AI, the obligations for high-risk systems only apply from 2 December 2027, and for systems under Annex I from August 2028. Companies gain time, but they do not lose the obligation.
What can a company do today?
Keep its own AI register, deploy only tested models, define clear rules for use and let humans intervene at the decisive points. That is AI governance on a small scale, and it works regardless of when the legislator is finished.
Related: AI Governance · EU AI Act
Recent Posts
Recent Comments