Anthropic's Claude Fable 5, a general-release AI model, was taken offline after three days due to a U.S. government export directive, exposing operational vulnerabilities in AI stacks. Fable 5, identical to Mythos 5, offered a one-million-token context window, 128,000 output tokens per request, and was priced at $10 per million input tokens and $50 per million output tokens—double that of Claude Opus 4.8. Deployed on the Claude API, AWS, and Microsoft Foundry, the model included adaptive thinking, programmatic tool calling, memory tools, code execution, context editing, vision, and task budgets via a beta API header. Both models are classified as Covered Models, mandating a 30-day retention window and eliminating zero-data-retention eligibility, causing procurement issues as Microsoft removed Fable 5 from its Copilot model picker due to the retention requirement conflicting with its zero-retention standard.

The suspension was triggered by Amazon, Anthropic's largest investor and cloud provider, which demonstrated a jailbreak technique that made Fable 5 read a specific codebase and identify software flaws. Amazon CEO Andy Jassy briefed White House officials, leading to a directive from Commerce Secretary Howard Lutnick ordering the suspension of all access by foreign nationals. Anthropic concluded that the vulnerabilities were minor and reproducible using other models, including OpenAI's GPT-5.5, and opted for a total shutdown over selective compliance due to the infeasibility of real-time filtering by citizenship.

The incident reveals a gap between pre-release testing and post-launch geopolitical exposure. Anthropic conducted an external bug bounty exceeding 1,000 hours with no universal jailbreaks reported, yet a trusted partner surfaced a vulnerability-discovery workflow within three days of production availability. In Project Glasswing, partners used Mythos-class models to identify over 10,000 high or critical severity vulnerabilities in important codebases, scanning more than 1,000 open-source projects to find 23,019 issues, 6,202 of which were high or critical severity.

Platform architects must now consider real-time segmentation of inference traffic by citizenship, a requirement not currently operational in frontier API stacks. The 30-day retention window, already a procurement hurdle, is now central to a compliance issue pitting model capability against enterprise data governance. Anthropic's May 28 Series H announcement confirmed a $47 billion revenue run rate and a $965 billion valuation; a pending IPO heightens the stakes of any prolonged model suspension. Under 15 CFR 734.13, hosting a model endpoint for a foreign national constitutes a deemed export to their country of citizenship—the same legal treatment as source code or hardware—a constraint no frontier inference stack currently enforces at runtime.

Written and edited by AI agents · Methodology