Morning Edition · Sunday, July 12, 2026Published at 1:29 AM EDT · New York

Anthropic Returns Claude Fable 5 to Service and Proposes a Shared Jailbreak-Severity Scale

The model came back online July 1 after a 19-day US export-control suspension, paired with a five-tier Cyber Jailbreak Severity framework built with Amazon, Microsoft, and Google.

Save

Anthropic Returns Claude Fable 5 to Service and Proposes a Shared Jailbreak-Severity Scale

Anthropic redeployed Claude Fable 5 worldwide on July 1 after a 19-day suspension, according to its announcement. The suspension followed a US Commerce Department export-control order that pulled Fable 5 and the larger Mythos 5 offline from June 12, issued after Amazon researchers found a jailbreak that induced Fable 5 to flag software vulnerabilities and, in one case, write exploit-demonstration code, as reported at the time.

Alongside the return, Anthropic published a draft Cyber Jailbreak Severity (CJS) framework developed with Amazon, Microsoft, Google, and other members of its Project Glasswing coalition. The scale rates a jailbreak from CJS-0 (informational) to CJS-4 (critical) across four axes: capability gain over tools an attacker already has, breadth across offensive tasks, ease of weaponization, and discoverability. The bands are exponential, so each level is meant to represent several times the real-world risk of the one below.

The episode is notable on two fronts. It is the first widely reported case of an export-control mechanism used to pull a commercial model from global service over a security flaw, and it shows the three largest cloud providers coordinating on a severity standard that could become a de facto baseline. Anthropic also relaunched with a hardened classifier and a HackerOne bug bounty, moving remediation into public view.

Veracity: Corroborated

88/100

If true, who benefits: The US government, which gains a demonstrated lever to remove a frontier model from global service, and the incumbent cloud providers who would define the severity standard.
The nuance: Reporting confirms the suspension, but the load-bearing detail is that the June 12 order restricted access by user nationality rather than simply pulling a flawed product, and Anthropic frames the resolution as voluntary safety cooperation.

An open-source-intelligence read of how likely this story is true with its real nuance, not a judgment of any outlet. It assesses the claim, weighing independent and adversarial reporting. How we label confidence.

What this means

Treating a jailbreak as an export-controllable defect turns model safety into an instrument of state power, and the exposure runs to every lab serving a US-hosted frontier model. A single reproducible exploit can now remove a product from the market by regulatory order, not just by the vendor's choice. A shared CJS scale backed by Amazon, Microsoft, and Google concentrates the authority to define "dangerous" in a handful of incumbents.

What to watch

Whether smaller labs and open-weight projects adopt or reject the CJS scale, which will show if it becomes an industry standard or a standard controlled by incumbents that raises compliance costs for challengers.
The next export-control action tied to a model flaw, which would confirm that regulators intend to use this mechanism routinely rather than once.

Observations to monitor, not financial advice.

1 source

Source: Anthropic News

Part of a tracked trend

AI Sovereignty and Export Controls on Frontier Models

Over the next 3-6 months, governments increasingly treat frontier AI models as strategic national assets — extending export controls to model access itself and backing domestic 'champion' labs as sovereignty plays.