This was the week the politics of AI stopped being theoretical. A single government order knocked two frontier models offline, and the aftershocks ran through Europe, the chip market, and every enterprise now betting its workflows on AI agents. Here are the ten stories that mattered most, counting up to the one that defined the week.
10. The job question got louder and more contradictory
Microsoft CEO Satya Nadella warned that a handful of frontier models could “hollow out” entire industries, the way globalization gutted manufacturing towns. At the same time TechCrunch described the AI layoff wave becoming a “powder keg” as a small group of insiders gets rich. The contradiction: UK AI hiring jumped 61%, mostly for people to babysit the bots, not build them.
9. The public is not buying the hype
A fresh Pew survey landed hard: 63% of Americans think AI is moving too quickly and only 16% expect it to have a positive impact on society. The twist is that unease is rising even as use climbs, with chatbot adoption now near half the population. People are leaning on these tools and distrusting them at the same time.
8. Governments started drawing hard lines
Norway imposed a near-total ban on AI in elementary schools, the UK pushed ahead with face-scanning age checks for asylum-seekers despite knowing the tech is flawed, and Bernie Sanders floated a $7 trillion plan to give Americans a stake in the AI industry. Education, surveillance, and economics all at once. The regulatory appetite is widening, and the biggest firms will resist every piece of it.
7. Physical AI consolidated while the money flowed
Hyundai took full control of Boston Dynamics, buying out SoftBank’s stake for $325 million. Nvidia showed off AI coding agents that autonomously direct robot training, even teaching robots to install GPUs. World-model startup Odyssey raised at a $1.45 billion valuation backed by Amazon, a sign that physical AI is where the next land grab is happening.
6. The agent security reckoning arrived
The same agent boom everyone is racing into is full of holes. Researchers found roughly 7,000 Langflow servers under active attack, with LangGraph and LangChain sharing the same exploitable bug class that turns ordinary flaws into remote code execution. Parallel disclosures hit Microsoft 365 Copilot and LiteLLM. The common thread is AI systems that trust external input without a boundary.
5. Enterprise software went fully agentic
This was the week agents stopped being demos and got embedded in the tools teams already use. Adobe pushed “creative agents” across Premiere, Photoshop, and Illustrator, GitLab 19 extended agents into merge requests and supply-chain scanning, and Salesforce agreed to buy customer-service AI firm Fin for $3.6 billion. The shift is from one-off generation to agents that run multi-step workflows.
4. The compute and chip race intensified
The infrastructure buildout kept compounding. Amazon is reportedly in talks to sell its in-house AI chips to outside data centers, a move Andy Jassy frames as a $50 billion opportunity and a direct shot at Nvidia. Nvidia itself moved to raise over $25 billion in its first bond sale since 2021, while startups like Tensordyne bet on logarithmic math to undercut GPUs. Everyone is racing to own the layer underneath the models.
3. Chinese open models closed the gap
Z.ai released GLM-5.2, a 753-billion-parameter MIT-licensed model that reportedly beats GPT-5.5 on coding benchmarks for one-sixth the cost, which Simon Willison calls probably the most powerful text-only open-weights LLM. The mood spilled into widely shared arguments that there is minimal downside to switching to open models. When open weights are this good and this cheap, the premium subscription starts to look optional.
2. The sovereign-AI surge went into overdrive
Washington’s ability to flip a switch turned “sovereign AI” from a slogan into a strategy. At the G7, world leaders said they want American AI but not an American off-switch, and the EU’s digital-sovereignty push surged despite US objections. Open foundation models pitched explicitly for sovereignty, like Apertus, suddenly read as infrastructure rather than idealism.
1. Washington pulled Anthropic’s models offline
The story that shaped everything else: a White House export-control order forced Anthropic to pull Fable 5 and the Mythos cybersecurity model, cutting access even for its own staff. Reporting then revealed the alarming “jailbreak” behind it was really just asking the model to “fix this code”, and dozens of security experts signed an open letter calling the ban counterproductive. Weeks later no one can say exactly what rule was broken, and the restriction may even be helping Anthropic’s brand.
The week’s lesson is simple. The hardest constraints on AI right now are not the models. They are the politics, the supply chains, and the trust around them.