For about three weeks this June, two of Anthropic's most capable models simply were not there. Not slow, not rate-limited. Gone. If you had Claude Fable 5 open in a tab, it stopped answering. That is a strange thing to watch happen to a product millions of people use every day, so it is worth going through what actually took place, in order.
The launch
Anthropic shipped Claude Fable 5 and Claude Mythos 5 on June 9, 2026. They are the same underlying model. Fable 5 is the generally available version, wrapped in the strongest safeguards Anthropic has ever put on a model. Mythos 5 has fewer of those guardrails and was only handed to a small group of vetted partners under a program called Project Glasswing, mostly for defensive cybersecurity work.
Three days after launch, everything stopped.
Why the government stepped in
On Friday, June 12, the US government applied export controls to both models. The order restricted access to foreign nationals, wherever they happened to be, including inside the United States. Anthropic had no reliable way to check nationality in real time, so it did the only thing it could do: it suspended access for every user, everywhere.
The trigger was a report from Amazon researchers. They had found a way to prompt Fable 5 so it would identify software vulnerabilities, and in one case, produce code that demonstrated how to exploit one of them. That sounds alarming until you read the rest of Anthropic's own account. When they tested the same prompt against other models — Claude Opus 4.8, GPT-5.5, Kimi K2.7 — those models found the same vulnerabilities too. Every model they tried, including much smaller ones like Claude Haiku 4.5, could reproduce the exploit demonstration once the vulnerability was named. So the finding was real, but it was not a uniquely dangerous Fable 5 capability. It was a gap in how the safety classifier drew its line.
What changed before it came back
Anthropic spent the suspension building a new classifier aimed squarely at the technique in the Amazon report. It now catches that specific pattern more than 99% of the time. Researchers from the Commerce Department's Center for AI Standards and Innovation reviewed both the old and new safeguards and signed off on the new ones as, in their words, extraordinarily strong. The trade-off is honest: the tighter classifier also blocks a few more ordinary coding and debugging questions than before. Anthropic says it will keep tuning that balance down over time.
On June 26, the government approved renewed access to Mythos 5 for more than 100 vetted US organizations. On June 30, the export controls on both models were lifted entirely. Fable 5 came back to everyone on July 1 — Claude.ai, the API, Claude Code, Claude Cowork, all of it.
The part that matters beyond Anthropic
The most useful thing to come out of this whole episode is not really about Fable 5. It is that there is still no shared, cross-industry way to describe how serious an AI jailbreak actually is. Anthropic, working with Amazon, Microsoft, and Google, has proposed a scoring framework built on four questions: how much capability the jailbreak actually adds, how many different attacks it works for, how much effort it takes to weaponize, and how easy the technique is to find in the first place. It is not finished. But it is the first serious attempt at something like a CVSS score for jailbreaks, and if it sticks, every AI vendor will end up using some version of it.
Why this should matter to your engineering team, not just AI Twitter
If your product depends on a frontier model for anything customer-facing, this story is a preview of a risk you have probably not written down anywhere. A government order can take a model offline worldwide with zero notice and no appeal window. It happened here in under 72 hours. Teams that had hardcoded claude-fable-5 with no fallback path spent three weeks either degrading gracefully or not degrading at all.
This is the kind of failure mode that a dedicated engineering team plans for during architecture review, not after an outage. If you are building AI features into a product and want a team that thinks about vendor risk this way from day one, that is exactly the kind of judgment our engineers bring to client projects — see how we approach dedicated development teams for the fuller picture.
