Warum sie wirklich eine KI Bremse wollen | Wasner + Steinschaden

A strange week in AI: the people usually associated with pushing the field forward are now asking for a pause. Dario Amodei, Sam Altman and Elon Musk have all called for slowing AI development, and the timing makes the whole debate look upside down. In the White House, that plea apparently lands on deaf ears. BRICS countries, meanwhile, are not waiting around.

That shift is the backdrop for a new episode from Jakob Steinschaden and Clemens Wasner, who pull apart a pile of fresh warning signs. One is the Hugging Face incident involving 700 AI agents, plus an attack on an Austrian developer wiki. Together, they raise awkward questions about what OpenAI’s testing really catches, and what it misses.

Another thread runs through Anthropic’s threat-intelligence report, which documents five stopped cases of bioweapon-related research. The report also touches on drones and rockets, and on how prompts are evaluated. And then there is the odd-looking move away from Chain of Thought: models may work better when nobody can read along, which is great for performance and terrible for audits.

The policy picture is just as scrambled. Donald Trump is said to reject regulation and prefer a “High IQ President,” while people in Silicon Valley worry that frontier labs may end up writing their own rules. Open weights could get squeezed in the process. China is pushing open weights for BRICS countries, Mistral is trying to balance that reality as a neocloud, and Europe has the same incentive to care.

Then comes the corporate side. OpenAI has pushed its IPO to 2027, Anthropic is reportedly eyeing two trillion dollars, and the prospectus question is what liability risks would have to be spelled out. Add Stripe buying OpenRouter and Nvidia taking Hugging Face, and two key access points to open weights are now in U.S. hands. That is not a subtle trend.