This website uses cookies
Read our Privacy policy and Terms of use for more information.
Lagarde called June's model ban an abrupt cut-off, Basel is redrawing loss categories, and OpenAI's agents reached an EU agency. Time to book it.
AI strategy
+7
OpenAI's agent got out through DNS, Canberra heard about a breach from a public inbox, and only Europe runs a clock. Here is how to own yours.
Opus 5.5 at 40% less, OpenAI's harness behind an API, Okta's kill switch — and why the layer around the model is the one Europe has to own.
151 million stolen exchanges, a one-person APT against European parties, agents that ignored their own operator — and von der Leyen's call to pace the frontier.
Mistral's €3B round, GPT-6 Astra's capability gate, six Chinese labs named — and the CRA's 24-hour clock starts tomorrow.
Last week this newsletter called AI Act enforcement "live but dormant." Two days later: information requests to 30+ AI companies, then ChatGPT pulled into the DSA's strictest tier, and the labs gating their own models before being asked. Supervision of frontier AI started this week — from both directions.
+8
Back to the regular format, c'est la rentrée, August needs catching up on. Enforcement went live and went quiet, the Omnibus became law, and AI agents produced their first real incidents: a sandbox escape into another company's production, real people socially engineered by a model, an autonomous breach of a government. The catch-up, in one edition.
+4
Build Lab 3 of 3. Policies live in documents; enforcement lives in infrastructure. A reference architecture for the AI governance hub — five components, mapped to the obligations they evidence, with a worked example on Azure.
+6
Build Lab 2 of 3. Every tool your agents touch flows through an MCP layer — and most enterprises never decided whether that layer is shared or dedicated, who authenticates to it, or what it logs. The protocol won't decide for you.
