AI Governance Daily, 11 May 2026

Here's the headline most outlets buried last week: the Trump administration ordered federal agencies to phase out Anthropic over the Mythos safety dispute. Then, within days, the same administration started drafting a plan to bring Anthropic back in.

That isn't a flip-flop story. It's the story.

AI procurement is now a live wire of federal policy. The Mythos model went from "supply chain risk" to "asset we need in government" in roughly a week. If you're an SME exec waiting for the rules to settle before you make a vendor decision, I have to be blunt: you're going to be waiting forever, and you'll lose the window.

The rules aren't being written ahead of time. They're being written on top of you, in real time, by people reacting to last Tuesday's model release. Plan accordingly.

🇪🇺 Europe & Regulation

• EU Omnibus clarifies who actually supervises what. The AI Office gets GPAI supervision, but national authorities keep competence over law enforcement, border, judicial, and financial use cases. Translation: if you're a deployer in a regulated sector, your regulator is still your regulator. Don't assume Brussels is your only counterparty. Read
• CNIL puts connected glasses on notice. Ray-Ban Meta and the smart-glasses-with-camera category get a direct warning on privacy and societal risk. If your workforce is starting to wear these into client meetings, that's now a data protection issue you own. Read

🇺🇸 US

• Trump orders federal agencies to phase out Anthropic. Six-month Pentagon window, triggered by the Mythos safety dispute and Anthropic's Project Glasswing limits. The signal: federal procurement is now a weapon in the AI safety debate. Read
• Same administration reverses within days. The White House is drafting guidance to bypass the supply chain risk designation and onboard Mythos. Whiplash = the new operating tempo. If your vendor strategy can't survive a week-long policy U-turn, it isn't a strategy. Read
• CAISI signs pre-deployment evals with Google DeepMind, Microsoft, and xAI. Stacked on top of the existing Anthropic and OpenAI deals. The US government now evaluates every major frontier model before public release. Quiet, multi-vendor, and a bigger structural shift than the Anthropic drama. Read
• White House preparing an FDA-style AI security executive order. Hassett's framing: "released to the wild after proven safe, just like an FDA drug." Pre-market approval for frontier models = the direction of travel. Read

🏢 Enterprise & Operating Model

• OpenAI publishes its "how enterprises are scaling AI" playbook. Trust, governance, workflow design, quality at scale. It's vendor-side framing, so read it with that lens, but the workflow-design point is the one I'd take seriously. Most enterprises are still bolting AI onto broken processes and calling it transformation. Read

🧠 Voices worth 5 minutes

• Azeem Azhar, Exponential View #573: "Are the AI labs building for an intelligence explosion?" Mythos preview, fusion economics, jobs, personhood. If you only read one thing this weekend, read this. It's the right altitude for board-level conversations. Read

⚠️ Watchlist

• Trump AI security executive order: expected. Hassett comparing it to FDA drug approval.
• Council and Parliament formal adoption of EU Omnibus deal: before 2 August 2026.
• Art. 50(2) provider transparency now lands 2 December 2026 (was 2 August 2026 before Omnibus).
• High-risk obligations (Annex III) now 2 December 2027.
• CAISI's expanding multi-vendor pre-deployment regime: watch for the first published evaluation findings.
• CNIL follow-up guidance on connected glasses in workplace settings.

---

The rules are being written on top of you, not ahead of you. If your AI governance plan assumes a stable rulebook, rewrite the plan, not the rulebook.