The United States government has directed Anthropic to prevent all foreign nationals from accessing its Fable 5 and Mythos 5 AI models — a move that has effectively forced the company to suspend both systems for users worldwide. The directive, first reported by BleepingComputer, marks a notable escalation in how Washington approaches the governance of advanced AI capabilities, and raises pointed questions about proportionality, effectiveness, and the precedent it sets for the global AI landscape.
A Global Ban Born From a National Restriction
The crux of the issue is practical. Because Anthropic has no reliable mechanism to verify the nationality of every user interacting with its models, the company concluded it had no choice but to halt access to Fable 5 and Mythos 5 entirely. A restriction nominally aimed at foreign nationals has thus produced a blanket worldwide suspension, cutting off legitimate users across every geography, including those in allied nations.
The order reportedly stems from a specific security concern: a jailbreak technique that could be used to extract harmful capabilities from the models. Anthropic, however, has pushed back on the government's reasoning. The company says the cited jailbreak is narrow in scope and does not unlock uniquely dangerous outputs. It further contends that comparable capabilities are already widely available through other sources — calling into question whether restricting its models achieves any meaningful security gain.
Proportionality, Effectiveness, and Precedent
The incident crystallises three tensions at the heart of modern AI governance.
Proportionality. When a targeted restriction produces a global shutdown affecting millions of uninvolved users, the cost-benefit calculus deserves scrutiny. The inability to enforce nationality-based access controls means a total suspension becomes the default — a blunt instrument that punishes the broader user base for a narrowly defined risk.
Effectiveness. If the capabilities the government seeks to restrict are, as Anthropic maintains, obtainable through other means, the order may achieve little beyond reducing access to one particular provider. Security measures that can be trivially circumvented risk being symbolic rather than substantive.
Precedent. Perhaps most consequentially, this action signals a shift toward applying export-control-style restrictions to the outputs of AI models, not merely the hardware and software used to build them. Historically, U.S. export controls have targeted chips, fabrication equipment, and source code. Extending similar logic to model access — particularly in a way that cannot be granularly enforced — could accelerate a fragmentation of global AI access along national lines.
Wider Implications for the AI Ecosystem
The episode is likely to reverberate across the industry. Other companies developing frontier models will be watching closely to assess whether similar orders could target their products. International regulators, meanwhile, may treat this as justification for building their own domestic AI frameworks, potentially producing a patchwork of divergent governance regimes.
For the open-source community and developers outside the United States, the case strengthens the argument for AI systems not subject to single-jurisdiction government orders. Open-source alternatives and models hosted in non-U.S. jurisdictions could gain momentum as users and organisations seek to insulate themselves from unpredictable access restrictions.
As of this writing, it remains unclear how long Fable 5 and Mythos 5 will stay offline, whether the U.S. government intends to refine its approach with more targeted controls, or how other nations will respond to the precedent. What is clear is that the boundary between national security policy and global AI access is becoming increasingly contested — and the consequences are being felt well beyond any single border.
美國政府已指示 Anthropic 阻止所有外國國民使用其 Fable 5 和 Mythos 5 人工智能模型——此舉實質上迫使該公司暫停向全球用戶提供這兩款系統。據 BleepingComputer 首先報導，這項指令標誌著華盛頓對先進人工智能能力治理方式的顯著升級，並就其合理性、有效性及其為全球人工智能領域所立下的先例提出了尖銳問題。
源於國家限制的全球禁令
問題的核心在於實際操作。由於 Anthropic 沒有可靠的機制來驗證每個與其模型互動的用戶的國籍，公司得出結論，別無選擇只能完全停止 Fable 5 和 Mythos 5 的訪問權限。一項名義上針對外國國民的限制，因此導致了全球性的全面暫停，切斷了各地區合法用戶的使用途徑，包括那些來自盟國的用戶。
據報導，該命令源於一項具體的安全隱憂：一種可用於從模型中提取有害能力的 jailbreak 技術。然而，Anthropic 反駁了政府的理據。該公司表示，所指的 jailbreak 範圍有限，並不能解鎖獨特危險的輸出內容。它進一步主張，類似的能力已通過其他來源廣泛獲得——這質疑了限制其模型能否帶來任何有意義的安全增益。
合理性、有效性及先例
此事件凸顯了現代人工智能治理核心的三大矛盾。
合理性。 當一項有針對性的限制導致影響數百萬無關用戶的全球停機時，其成本效益分析值得仔細審視。無法實施基於國籍的訪問控制，意味著全面暫停成為預設方案——這是一個鈍器，因一個範圍明確的風險而懲罰更廣泛的用戶群體。
有效性。 如果政府尋求限制的能力，正如 Anthropic 所堅稱的那樣，可以通過其他方式獲得，那麼該命令除了減少對某一特定供應商的訪問外，可能收效甚微。那些可被輕易規避的安全措施，其風險在於僅具象徵意義而非實質作用。
先例。 或許影響最深遠的是，此舉預示著一種轉變，即將出口管制式的限制應用於人工智能模型的輸出內容，而不僅僅是用於構建它們的硬件和軟件。歷史上，美國的出口管制主要針對晶片、製造設備及 source code。將類似的邏輯擴展到模型訪問——尤其是在無法細粒度執行的情況下——可能加速全球人工智能訪問沿國家界限出現分裂。
對人工智能生態系統的更廣泛影響
此事件很可能在整個行業引起反響。其他開發前沿模型的公司將密切關注，以評估類似命令是否會針對其產品。與此同時，國際監管機構可能將此視為建立本國人工智能框架的理由，從而可能產生一系列相互分歧的治理體制。
對於開源社群及美國以外的開發者而言，此案例加強了論點，即人工智能系統應不受單一司法管轄區政府命令的約束。開源替代方案以及託管在非美國司法管轄區的模型，可能會因用戶和組織試圖將自身與不可預測的訪問限制隔絕開來而獲得發展勢頭。
截至撰寫本文時，Fable 5 和 Mythos 5 將會停機多長時間，美國政府是否打算以更有針對性的控制來調整其做法，或其他國家將如何回應此先例，目前仍不清楚。但清楚的是，國家安全政策與全球人工智能訪問之間的界限正變得日益有爭議——其後果正遠遠超出任何單一邊界而被感受到。