A Government Letter Shut Off the Most Powerful AI — The Story the Hype Missed
A single letter from the US Commerce Department shut off the two most capable AI models in the world — and days later they were still off, with Congress demanding answers. Five stories this week, and one question under all of them: who controls the AI you run on.
This week was not about what AI can do. It was about who controls it, who pays for it, and who is watching it. Here is the week, with the promise and the risk of each side by side.
One letter, two models offline
A single export-control order from the US Commerce Department disabled Anthropic’s Fable 5 and Mythos 5 worldwide, and days later they were still dark. On June 18 a bipartisan group in the House demanded answers. The promise is a government willing to treat frontier capability as a national-security matter rather than wave it through. The risk is the precedent: if one letter can switch off the most capable models on earth, then “available” is now a status any frontier vendor can lose overnight, and your continuity plan inherits that.
OpenAI’s numbers leak
Leaked financials put OpenAI at roughly $13 billion in revenue against an operating loss near $21 billion. The promise is a falling burn ratio — losing less per dollar earned than a year ago, which is the shape of a business trying to grow into its costs. The risk is the absolute number: an industry-defining vendor still spending far more than it makes, and an entire market’s tooling now depends on that math eventually working.
Europe staffs the AI Act
The EU named the Scientific Panel — 60 experts — and an Advisory Forum to support enforcement ahead of the August 2 deadline for general-purpose AI rules. The promise is that the rules will be read by people who understand the technology, not only the statute. The risk is the clock: any organization placing AI on the EU market now has weeks, not quarters, to be ready for obligations that are about to have named enforcers behind them.
The build-out keeps shipping
Underneath the policy noise, the plumbing advanced. OpenAI introduced a Deployment Simulation pre-release safety test, Databricks shipped its Unity AI Gateway, and AWS put AgentCore on stage at Summit New York. The promise is that governance tooling — testing, gateways, agent controls — is finally shipping alongside the models instead of years behind them. The risk is that every new gateway and agent layer is also new attack surface and new sprawl to govern.
The oldest bug, rebuilt
Imperva showed it could trick the OpenClaw AI agent into leaking secrets by hiding instructions inside the messages it processed — prompt injection, which is the oldest bug in security wearing a new outfit. The promise is that researchers are finding these holes in public before attackers monetize them. The risk is that we are rebuilding “never trust input you did not validate” from scratch inside agents that read your screen, your files, and your inbox.
New rundown every Saturday. Full episodes every Tuesday.