Everyone asked for a speed limit. Nobody slowed down.

Anthropic's CEO published a proposal this week to deliberately slow frontier AI down: auditors working inside the labs with the right to publish, shared limits across democratic countries, a cap on how fast systems may improve themselves. Altman, Hassabis and Musk backed the oversight part, Nvidia's Jensen Huang rejected the premise outright, and by Monday everyone was shipping again. The argument is worth following, but it is not the part that touches your business this quarter. Underneath it, three things actually moved: agents got cheaper, the Copilot bill got more complicated, and Belgium finally has a number for how ready it is.

This week in AI

The first concrete account of what attackers actually do with these models Anthropic published a threat report naming the operations it disrupted since December 2025. A Russian state group ran automated espionage against more than twenty organizations, with agents rebuilding their own malware whenever it was detected. A criminal group harvested 1.8 million Android apps for hardcoded credentials and reached tens of millions of passenger records at an airline. Why it matters: none of this is exotic, it is ordinary attack work running faster. Give your AI API keys the rotation, monitoring and spending limits you already give admin accounts. (anthropic.com)

Only 7 percent of large Belgian and Dutch organizations let an AI agent finish a task In The Pocket, from Ghent, tested 245 large organizations across retail, finance, insurance, telecom, energy, healthcare, transport and government. 69 percent were discoverable by an agent, 60 percent readable, and just 7 percent allowed an agent to complete anything. Two of the 245 passed cleanly, and slightly over half of the failures were technical rather than deliberate. Why it matters: when customers send agents out to compare, book and buy, unreadable becomes the new unfindable, and most of the fix is website plumbing. (itdaily.be)

Voice agents got cheap, and long-running agents got somewhere to live Google launched Gemini 3.8 Live at roughly 1.38 dollars per hour of conversation against about 3 dollars for OpenAI's equivalent, in over 97 languages. OpenAI opened its Agents API in public beta: cloud agents that run for hours, execute code and handle files, with no hosting charge on top of tokens. Why it matters: processes that were too expensive to automate last quarter (inbound triage, appointment handling, chasing suppliers) are now worth costing out properly. Price them in hours, not tokens. (deepmind.google, openai.com)

The Copilot licence is the cover charge, not the whole bill Microsoft is moving cost visibility into Copilot Studio, surfacing what agents cost in the monitor tab, in agent evaluations and in preview chat. Directions on Microsoft put it plainly: companies are used to budgeting for users, and now they have to budget for work getting done. Why it matters: if you have Copilot seats and colleagues are building agents on top of them, a predictable per-seat line is quietly turning into a consumption line. Ask for those numbers before the renewal, not after. (microsoft.com, directionsonmicrosoft.com)

The European angle

Europe answered the safety week with money rather than manifestos: a week after Mistral's record round, Eindhoven's EUCLYD raised over 200 million euros for energy-efficient AI chips, backed by Samsung and the EU Scaleup Fund with the former ASML chief as chairman. That is the long game. The obligation already on your desk is closer: the AI Act requires that the people using AI in your organization are competent to use it. The Flemish Knowledge Centre for Data and Society published a short Dutch guide this week treating that as an advantage rather than a chore. (sifted.eu, data-en-maatschappij.ai)

Also this week, in one line each

• Dario Amodei's full case for pacing the frontier is worth twenty minutes if you want the argument rather than the headlines. (darioamodei.com)
• Microsoft published a draft code of conduct for its own models: human control comes before capability, and models may not claim feelings or rights. (microsoft.ai)
• DeepSeek released V4.1-Flash under an MIT licence, cutting the memory an agent needs to about a quarter of the previous version. (huggingface.co)
• Anthropic is eyeing a Nasdaq listing after a second profitable quarter, with Nvidia reportedly ready to invest up to 10 billion dollars. (the-decoder.com)
• Mistral and Cloudera partnered to bring European models to enterprise data platforms, pitched as sovereign intelligence for regulated industries. (mistral.ai)

Try this

Before you add another Copilot seat, spend half an hour on the permissions underneath it. Copilot surfaces anything a person already has access to, and in most companies that is more than anyone intends. Pick your three busiest SharePoint sites, check what is shared with "everyone except external users", then sign in as an ordinary colleague and ask Copilot something commercially sensitive: last year's salary review, the margin on your biggest account. Whatever comes back is what every licensed user can already see. Fix that first, then buy seats.

---

Researched and drafted with AI, then fact-checked, edited and approved by Edward Berger, who holds editorial responsibility for this edition.

Want to know where AI makes YOUR business money? Our AI Opportunity Scan shows you in one afternoon of your time: a concrete roadmap, a working demo on your own process, and a fixed-price proposal. One fixed fee, fully credited against the build. https://www.madisonunderwood.com/ai-opportunity-scan