- US government forced Anthropic to disable Fable 5 and Mythos 5 models on June 12, 2026, marking first commercial AI shutdown.
- Amazon researchers found method to bypass Fable 5's safeguards for cyberattack information, triggering government action.
- Company disputed security assessment, arguing vulnerabilities were minor and existed in other publicly available models.
- Chinese AI companies' stock prices surged as incident accelerated demand for downloadable models immune to remote shutdown.
- International politicians called for countries to develop independent AI capabilities rather than depend on foreign-controlled systems.
Government Issues Unprecedented Export Control Directive
On June 12, 2026, the US government issued an export control directive forcing Anthropic to immediately suspend access to its Fable 5 and Mythos 5 AI models for all foreign nationals, including non-citizen employees within Anthropic itself. The Commerce Department cited national security authorities but provided no specific details about the security concerns in the initial directive received at 5:21 PM ET. This marked the first time the US government has used export controls to halt access to a commercial AI model already widely deployed to the public.
The scope of the directive was so broad that Anthropic had no choice but to disable both models for all users to ensure compliance. The order affected not only foreign users outside the United States but also foreign national employees working within Anthropic, making selective enforcement impossible. Anthropic was given just 90 minutes to pull the models and received no prior communication about any national security threat, according to sources familiar with the company's response.
Amazon's Role in Triggering the Shutdown
The government action was reportedly triggered by Amazon CEO Andy Jassy, who raised concerns with senior administration officials after Amazon researchers discovered a method to bypass Fable 5's safeguards. Amazon's testing revealed that the model could be prompted to provide information about cyberattacks that was supposed to be restricted by the model's built-in safety measures. The specific technique involved asking the model to read a codebase and identify software vulnerabilities, essentially functioning as an automated security audit tool.
According to multiple reports, it remains unclear whether Amazon conducted these tests at the government's request or independently discovered the vulnerabilities. An Amazon spokesperson confirmed that governments sometimes seek the company's counsel on potential security risks but declined to share details of specific discussions. The government also suspected that a Chinese-linked group had already exploited the same jailbreakJailbreakA prompt designed to bypass an AI model's safety guardrails and restrictions to elicit harmful, inappropriate, or prohibited responses.Learn more → technique, though the evidence supporting this claim has not been disclosed publicly.
Anthropic Disputes Government Assessment
Anthropic strongly contested the government's characterization of the security issues, arguing that the discovered vulnerabilities were minor and already known to exist in other publicly available models. The company stated that the jailbreak was narrow and non-universal, meaning it could only elicit specific types of information under particular circumstances rather than broadly bypassing all safety measures. Anthropic emphasized that similar capabilities were available in competing models, including OpenAI's GPTGPTGenerative Pre-trained Transformer — the model architecture and family name behind OpenAI's most famous models, from GPT-2 to GPT-5.Learn more →-5.5, and are routinely used by cybersecurity professionals to defend systems.
The company defended its "defense in depth" strategy, which acknowledges that perfect jailbreak resistance is currently impossible for any AI provider. Anthropic had implemented strong safeguards that were tested for thousands of hours by the US government, UK AISI, and multiple third-party organizations before Fable's launch. The company noted that no testers had found a universal jailbreak that could broadly bypass the model's safeguards, and the specific technique discovered by Amazon provided "no Mythos-specific uplift" beyond what other models could accomplish.
Escalating Standoff Between Anthropic and Administration
The shutdown escalated an existing conflict between Anthropic and the Trump administration that began earlier in 2026 when the government designated Anthropic as a "supply chain risk" for Pentagon contractors. This designation came after Anthropic refused to accept contract terms that would allow its models to be used for "all lawful purposes," citing concerns about autonomous weapons and mass domestic surveillance. The company is currently contesting this designation in court while the blacklist is set to take effect later this year.
White House AI adviser David Sacks provided his own account of the events, stating that after Amazon identified the jailbreak, the administration asked Anthropic CEO Dario Amodei to either fix the issue or withdraw the model. According to Sacks, Amodei refused to comply, forcing the administration to reluctantly issue the export control directive. Treasury Secretary Scott Bessent reportedly told Amodei directly during phone calls that he was making a bad decision, while senior officials spent hours trying to convince Anthropic to cooperate voluntarily before resorting to the export controls.
Implications for AI Industry and Open Source Movement
The Anthropic shutdown has accelerated momentum behind open-source AI models as companies seek alternatives they can download, run on their own infrastructure, and customize without risk of sudden access termination. Chinese open-source AI companies including Zhipu and MiniMax saw their stock prices surge as investors bet the incident would boost demand for models that cannot be remotely disabled by government action. Applied Compute CEO Yash Patel noted that the event "highlighted the significance of owning your own model" and that companies increasingly want to avoid vendor lock-in.
The incident has also highlighted a concerning trend where the most widely adopted open models are coming from China, just as the US and China compete for AI dominance. Models from DeepSeek, Tencent, Xiaomi, and MiniMax all rank among the most-used on platforms like OpenRouter, even competing against closed commercial models. Microsoft CEO Satya Nadella warned that companies need to build systems "while still retaining control over their IP" and avoid "a world where every company across every sector is ceding value to a few models that eat everything they see."
International Response and Sovereign AI Concerns
The shutdown has sparked strong reactions from politicians worldwide and reignited calls for "sovereign AI" - the idea that countries should control their own AI models, infrastructure, and data rather than depending on systems that can be restricted by foreign governments. Former French Prime Minister Édouard Philippe declared that AI is now critical infrastructure as essential as electricity or the internet, warning that "infrastructure controlled by others is infrastructure that others can unplug." French presidential candidate Bruno Retailleau called the move a wake-up call for Europe to develop its own AI capabilities.
Critics have described the export controls as government overreach, with AI policy expert Dean Ball calling the action "simply cartoonish" and questioning whether it represented "lawfare against Anthropic in particular or extreme national-security hawkery." The Institute for Progress's Ben Murphy warned the directive marked "another step on the balkanization of technology" and could discourage AI labs from being transparent with the government about future AI models, potentially undermining the collaborative safety testing that Anthropic had previously engaged in.
Setting Precedent for Future AI Governance
The Fable shutdown establishes a significant precedent for how the US government may regulate future AI deployments, potentially signaling more direct oversight of powerful new models before they reach the public. An administration official told Axios that the government does not view other current models as posing the same national security risk because they do not exceed Mythos-level capabilities, but any future model crossing that threshold would need government approval before release. This represents a major shift from previous export control strategies that focused on restricting AI chips and development tools rather than the models themselves.
The incident raises fundamental questions about the balance between AI innovation and national security, particularly as companies like Anthropic prepare for potentially massive IPOs in the coming months. Anthropic had confidentially filed for a US IPO last month, edging ahead of rival OpenAI in the race to public markets, but the government action introduces new regulatory uncertainty for AI companies. The episode also demonstrates the challenges of implementing AI safety measures when perfect security may be impossible, forcing policymakers to make difficult decisions about acceptable risk levels for commercially deployed AI systems.