The 19-Day Blackout That Changed How to Think About AI Risk

On June 12, a US government order cut off global access to Anthropic's Fable 5 — one of the most capable AI models available — giving Anthropic just 90 minutes to comply. Businesses and developers who depended on it woke up to broken workflows. The cause: Amazon researchers had found a technique that could trick Fable 5 into identifying software vulnerabilities and writing exploit code. The government treated that as a national security risk and pulled the model under export-control rules. It proved that one research report can cut off global access.

For 19 days, teams scrambled. Anthropic automatically routed about 5% of sessions to a fallback — Claude Opus 4.8 — but most users had no backup plan. Switching models is not a quick fix: it means rewriting prompts, recalibrating workflows, and accepting different output quality. Most teams had never tested what would happen if their primary model disappeared. The teams that had done that work in advance fared far better.

Access returned on July 1, after Anthropic agreed to three conditions: detect security risks proactively, coordinate with the government on future model releases, and report malicious use. Anthropic also built a new safety filter that blocks the reported jailbreak technique in more than 99% of cases. The bigger lesson is about power: frontier AI access is now conditional on government approval, and that changes the risk calculation for anyone building with these tools.

This is not a reason to stop using powerful AI. It is a reason to plan as if your main model could go offline next week — because now you know it can. Test a backup model today. Keep your prompts modular so you can swap providers. Map which of your products or workflows would break first. The question is no longer whether AI access can be interrupted. The question is whether your team is ready.