The Government Switched Off the Most Advanced AI Model on the Market. It Stayed Dark for 18 Days.

On June 9, Anthropic launched Claude Fable 5 and Claude Mythos 5, the most capable AI models it has ever shipped. On June 12, the US government applied export controls to both, and Anthropic switched them off - for every customer, on every surface. Not degraded, not rate-limited: off. The models disappeared from Anthropic's own platform and from AWS, Google Cloud, and Microsoft Foundry alike, and they stayed dark until the controls were lifted on June 30. It was the first time a government order has removed a commercial frontier model from the market, and it lasted eighteen days.

0

days live before the order

0

days dark under export controls

The directive took effect immediately on June 12. Global availability returned July 1, the day after the controls were lifted.

What actually happened

The trigger was a jailbreak. Researchers at Amazon found a prompting technique that bypassed Fable 5's safeguards and got the model to identify software vulnerabilities - and, in at least one case, to produce demonstration exploit code. The government's response was an export-control directive requiring Anthropic to restrict the models from foreign nationals, inside or outside the United States, effective immediately. Anthropic had no way to verify nationality in real time, so it did the only compliant thing available: it suspended access for everyone.

The detail that should reframe the whole episode: Anthropic's own investigation found the same technique worked on weaker models, including Claude Opus 4.8, OpenAI's GPT-5.5, and Kimi K2.7. The capability that triggered the order was not unique to the model that got switched off. What was unique was the enforcement action - and the enterprises caught inside it.

The outage nobody had a runbook for

For teams that had wired one provider's SDK straight into production, June 12 was a total outage with no failover. The next request had nowhere to go. Switching to another model meant a code change, a review, a deploy, and a release window - all executed under incident pressure, while the product was down, for an event no risk register had a line for. Availability did not fail the way availability usually fails. The vendor was healthy, the API was up, the status page was green. The model was simply no longer allowed to answer.

The failure was not degraded performance. It was total, immediate, and legally mandated - and it arrived with zero days of notice.

The restoration came with its own governance lesson. Anthropic trained a new safety classifier that blocks the specific technique from the Amazon report, and it is effective in more than 99 percent of attempts. The trade-off is false positives: some legitimate coding and debugging queries now get flagged and rerouted to a weaker model. Even the fix illustrates the point - the behavior of the dependency changed between the day it went dark and the day it came back, without any change on the customer's side.

0% of attempts blocked against the specific reported technique

050100

Cost of the classifier: some legitimate coding and debugging queries are now false-positived and rerouted to Opus 4.8, with the user notified.

The precedent is the story

The terms of the redeployment matter more than the outage. To bring the models back, Anthropic agreed to pre-release government access to future frontier models, proactive threat detection and reporting, rapid information sharing on significant jailbreaks, and work toward shared industry security standards. Whatever you think of those terms, they establish the mechanism: frontier model availability is now a regulatory variable, subject to change on government timelines, not vendor ones. This was also not 2026's only demonstration. OpenAI pulled a deployed model back over behavioral failures last year - covered in our post on the sycophancy rollback - and vendors revoke, deprecate, and regionally restrict models routinely. The causes differ; from inside your stack they all look identical. The dependency vanished.

Eighteen days later the models are back, better defended, and carrying new terms. The enterprises that felt nothing had a validated second path before June 12. The ones that went down are writing the runbook they should have had - for a failure mode that is now precedented, mechanized, and one directive away from happening again.