The first model recall in history — and how it ended

The story we told in June has found its ending — and a faster one than many expected. On 30 June the US government lifted the export controls; since 1 July, Claude Fable 5 is available worldwide again. What happened in the eighteen days in between is instructive.

The trigger was smaller than the alarm

Amazon researchers had found a method to bypass Fable’s safeguards and have the model identify software vulnerabilities. Anthropic’s retests showed that the same vulnerabilities were found by far weaker models too — down to Haiku 4.5 and competitor models. So the jailbreak did not unlock uniquely dangerous capabilities; it opened a grey area that was blocked out of caution anyway.

The response is technical and political at once

Anthropic trained a new safety classifier that blocks the reported technique in over 99% of cases (blocked requests now route automatically to Opus 4.8). At the same time, Amazon, Microsoft and Google are building an industry framework for rating jailbreak severity — four criteria: capability gain, breadth, weaponisability, discoverability. A kind of CVSS for AI models.

And the state now has a permanent seat at the table: pre-release access for authorities, disclosure duties on jailbreaks, joint safety research. What began as a crisis response is becoming standard process.

Also noted: the pattern is now the norm

On 9 July, OpenAI shipped its GPT-5.6 family — which had likewise launched on 26 June as a restricted preview under government conditions. The pattern from the Fable case is officially no longer an Anthropic episode; it is the new normal for frontier releases.

Our take

The incident confirms both theses from our June issue. First: model availability is a political risk — but one that can dissolve within weeks if provider and government are able to cooperate. Second: anyone who had a fallback architecture in June barely felt the outage; anyone who didn’t now has a very concrete business case for provider independence.

New is a third point: the jailbreak framework creates, for the first time, an objective yardstick for when a security incident justifies a shutdown — reducing exactly the arbitrariness that the June incident produced.

Thinking about provider independence and fallback architectures? Reach out to the contact below.