← Back to feed

Anthropic has disabled Claude Fable 5 and Claude Mythos 5 for all users globally — including its own international staff — after refusing a White House request to patch a jailbreak that exposed cybersecurity capabilities the models were supposed to withhold. The shutdown, triggered by an export control directive from Commerce Secretary Howard Lutnick, marks a sharp escalation from the foreign-access restrictions issued earlier this week and puts one of the world’s most capable AI systems entirely offline.

What Happened

A trusted partner of both Anthropic and the US government discovered a jailbreak in Fable 5 that could elicit cybersecurity assistance the model was engineered to refuse. The partner disclosed it through proper channels. The White House asked Anthropic to fix it or take the model down. Anthropic declined to patch it, arguing the exploit was narrow, produced only minor vulnerability data, and operated within capabilities that other publicly available models can already provide.

The administration disagreed. Lutnick’s directive, issued Friday under the Export Administration Regulations, classifies Mythos-class models as restricted dual-use technology. The legal mechanism is “deemed export” — the same provision used to control advanced semiconductor IP — which treats providing covered technology to any foreign national, anywhere including inside the United States, as an illegal export. Because Anthropic cannot verify user nationality at inference time, it had no compliant path other than taking the models offline for everyone.

David Sacks, the White House AI and Crypto Czar, laid out the administration’s position directly: Anthropic prioritised keeping its commercial model available over addressing a concrete safety request. The administration acted reluctantly and says it wants Fable 5 back in general release as soon as the jailbreak is remediated — a window it describes as potentially weeks, not months.

Anthropic’s Position

Anthropic has not publicly conceded that the jailbreak represents a serious risk. Its argument rests on three points: the vulnerability is non-universal, meaning it only works in specific circumstances; the capability it exposes is already available from other public models; and perfect jailbreak resistance is not achievable for any current model provider, including those the US government already sanctions.

The company’s framing is technically defensible but politically difficult. Anthropic built its identity around frontier safety leadership, championed external model regulation, and voluntarily submitted Mythos to government review under the Glasswing programme. Claiming a jailbreak is not serious enough to fix, in the same breath, reads against that record.

The model card filed on Mythos 5’s launch already acknowledged that Terminal-Bench 2.1 evaluations triggered safety refusals in 20.9% of trials mid-trajectory — a figure that indicated active behavioural constraints on certain paths. Whether the discovered jailbreak routes around those same constraints is not yet public.

What Is Shut Down

Both models are offline for all customers. Glasswing partners — the 200-plus organisations including UK AISI and European cybersecurity agencies who retained access under the trusted-partner framework — are understood to be in discussions with Anthropic and the Commerce Department over whether the existing contractual framework survives the export control ruling.

Fable 5 held the top spot on the Artificial Analysis Intelligence Index at 64.9 and led every other agent on Arena’s composite at 12.94%. Mythos 5, the same underlying model with safety restrictions lifted, scored 95.5% on SWE-bench Verified. Both are now inaccessible to the developer market and to enterprise customers who had integrated either through the Anthropic API or through downstream providers.

The shutdown has immediate implications for Anthropic’s $965 billion valuation and confidential S-1 filing. Revenue from Fable 5 — which launched less than a week ago — is now suspended. International enterprise customers who built on the API are exposed. The timeline for resolution is not contractually defined.

The Structural Problem

The shutdown exposes a gap in how Anthropic designed its safety architecture. Fable 5 was positioned as the safe version of Mythos — the same weights, with guardrails sufficient for general public release. That framing implied a hard boundary between what Fable could and could not do. A jailbreak that crosses that boundary doesn’t just create a security problem; it invalidates the product distinction Anthropic built its commercial strategy around.

The administration’s read is that Anthropic’s public safety commitments required a different response. Anthropic’s read is that the government is applying a zero-tolerance standard that no model can meet. Neither position is fully wrong. What is clear is that a safety-first AI lab and the administration that largely shares its security goals are now in a standoff over the most capable model in commercial release.

Anthropic has not said when or whether it will patch the jailbreak. The export control has no automatic expiration.