AI is now an interruptible supply chain | Monday, July 6, 2026

The story is that AI governance stopped being a white paper three weeks ago, when the US Commerce Department ordered Anthropic to switch off two of its most capable models and the company complied worldwide for nearly three weeks. That is the shift enterprise leaders should register.

For two years, AI governance meant voluntary principles and panel discussions. In the last month it acquired teeth: an export order that took a live model offline, an executive order routing frontier releases through federal review, and an intelligence-alliance warning that AI cyberattacks are months out. Access to the models you build on is now something a government can grant, delay, or revoke. Plan accordingly.

The proof point is Anthropic. On June 12, Commerce Secretary Howard Lutnick sent CEO Dario Amodei a legally binding export-control directive to cut off all access to Claude Fable 5 and Mythos 5 by any foreign national, anywhere, including foreign-national employees inside the US. Unable to filter users by nationality in real time across dozens of clouds, Anthropic shut the models off for everyone. The outage ran about 18 days until the order was lifted and the models returned July 1. It is the first time frontier models were switched off by regulatory order.

The machinery is thickening fast. A Trump executive order now asks labs to hand frontier models to federal agencies up to 30 days before public release for cybersecurity benchmarking, with a voluntary framework reportedly landing around August 1. OpenAI already launched GPT-5.6 behind a government-managed access list. Five labs are set to adopt a shared jailbreak severity scale. None of this is finalized, and most of it is nominally voluntary. But the direction is unmistakable: model access is becoming a governed, interruptible dependency.

The comforting reading of this week is that AI is finally being governed responsibly, with summits, frameworks, and severity scales. Look at the one action that actually had force and a different picture emerges. The Fable 5 shutdown was triggered by a jailbreak report, executed as a worldwide cutoff because nationality filtering is impossible at scale, and ran nearly three weeks. That is not a scalpel. It is a sledgehammer, and it landed on every legitimate enterprise user of the model regardless of where they sat or what they were building.

So the near-term risk for most enterprises is not over-regulation of frontier capability. It is unpredictable availability. A model you depend on can go dark for reasons that have nothing to do with you, on a timeline you do not control, decided by a letter you never see. Meanwhile the mundane exposure the Five Eyes flagged, unpatched systems and under-defended suppliers, gets none of the summit's attention. The governance that matters most to your operations is the governance nobody is convening a summit about.

We watched the Fable 5 episode closely, because it rehearses a risk most AI roadmaps ignore. Teams plan for model quality, latency, and cost. Almost nobody plans for a model simply becoming unavailable, not from an outage but from an order. When that happened last month, the enterprises that were fine were the ones with a second model already wired in and a tested path to switch. The ones that were not spent three weeks improvising.

The lesson is old and boring, which is why it is easy to skip. Resilience is unglamorous. It is a fallback model you rarely use, a failover you test on a quiet Tuesday, a contract clause about regulatory suspension that your legal team finds tedious. None of it shows up in a demo. All of it shows up the day your primary model goes dark.