# US Government Orders Anthropic to Disable Fable 5 and Mythos 5 Over Export Control Concerns
On June 12, 2026, the US government issued an emergency export control directive ordering Anthropic to immediately revoke access to its two most advanced AI models—Fable 5 and Mythos 5—from all users worldwide, citing national security grounds. The order bars access by any "foreign national," a definition broad enough to effectively take the models offline entirely, since Anthropic itself employs foreign nationals and cannot segregate users by citizenship across a global platform.
The decision came just three days after Anthropic launched Fable 5 free to millions of Pro, Max, and Enterprise customers in what was scheduled to be a two-week promotional rollout. The sudden suspension has left users with broken sessions, failed API requests, and lingering questions about whether this precedent signals a fundamental shift in how governments regulate frontier AI capabilities.
## What Are Fable 5 and Mythos 5?
Fable 5 and Mythos 5 represent Anthropic's latest model family, sharing the same underlying foundation but serving different security postures:
Both models represent significant capability jumps over Claude Opus 4.8, which remains unaffected by the order. The free Fable 5 rollout, which began June 9, was meant to run through June 22, giving Anthropic's broader user base early access before wider deployment.
## Background: National Security and AI Export Controls
The directive invokes national security authorities, placing Anthropic's models alongside semiconductors, cryptography, and military technology in the export control regime. This is not the first time the US government has restricted access to advanced AI systems—similar controls have been proposed for GPT-series models and other frontier systems—but it represents one of the most aggressive enforcement actions to date.
The regulatory framework draws from two overlapping authorities:
| Authority | Purpose | Application |
|-----------|---------|-------------|
| Export Administration Regulations (EAR) | Control US technology from reaching "denied persons" and foreign countries | Applies to AI model access across borders |
| International Traffic in Arms Regulations (ITAR) | Restrict military-related technology | Can be applied to dual-use AI systems |
| Entity List / Denied Parties | Target specific organizations and individuals | Enforcement mechanism |
The order's timing aligns with broader US government concern about AI capabilities reaching adversaries or proliferating unchecked to foreign entities, particularly those in China and other strategic competitors.
## The Alleged Jailbreak
According to Anthropic's disclosure, the government's directive stems from demonstration of a potential jailbreak affecting Fable 5. The company describes the method as "narrow and non-universal," requiring an attacker to:
1. Request the model to analyze a specific codebase
2. Ask it to identify software vulnerabilities
3. Potentially bypass safety guidelines through the analytical request
Anthropic's assessment is damaging to the government's case: the company says it reviewed the demonstration and found only "minor, already-known bugs" of the kind that other publicly available models—including OpenAI's GPT-5.5—can already discover without any exploit. "Our understanding is that one potential jailbreak was shared with the government," the company stated, implying that a single demo or report prompted the emergency order.
The company emphasizes that the capability in question—analyzing code for flaws—is precisely the functionality that security defenders rely on daily across the industry. Restricting it would undermine legitimate defensive security work.
## Anthropic's Response and Disagreement
Rather than quietly accept the directive, Anthropic filed a measured but firm public pushback:
The company promised more details within 24 hours and said it is working to restore access, suggesting it views the order as a misunderstanding that can be resolved through dialogue.
## The Practical Challenge: Defining "Foreign National"
The order's requirement to block "foreign nationals" from accessing Fable 5 creates a near-impossible compliance scenario:
Who is a foreign national?
Why this makes compliance effectively impossible:
Anthropic chose to disable the models for all users rather than attempt targeting, acknowledging that granular geo-blocking or identity verification at scale is not technically feasible without massive friction and privacy invasions.
## Industry Implications and Precedent Risk
This order sets several concerning precedents:
On Model Deployment: Frontier model developers may face emergency suspensions based on unverified vulnerability reports, creating perverse incentives to hide rather than disclose potential issues.
On Export Controls: The order extends US jurisdiction over software services globally, potentially fragmenting the internet and forcing companies to choose between US markets and international availability.
On Definitions of Risk: A single narrow jailbreak—one that requires specific knowledge of how to prompt the model and what codebases to analyze—is treated as grounds for wholesale model recall.
On Competitive Dynamics: If similar jailbreaks are found in GPT-5.5 or other systems but not acted upon with the same speed, this creates asymmetric regulation that favors some vendors.
## Recommendations for Organizations
For Enterprise Users:
For AI Teams:
For Security Organizations:
For Policymakers:
---
## HackWire Analysis
This order reveals a fundamental tension in frontier AI regulation: the gap between the government's security concerns and the practical enforcement mechanisms available.
The core problem is one of proportionality and process. Anthropic received a verbal directive at 5:21 PM on a Thursday evening with no advance notice and, by the company's account, no formal written evidence of a critical vulnerability. The jailbreak in question—asking the model to analyze code—is a legitimate use case. It's the kind of work security defenders do every single day. Other frontier models offer equivalent capabilities, yet they're not (publicly) facing equivalent treatment.
This sets a dangerous precedent. If any narrow, theoretically-reproducible jailbreak justifies emergency model recall, then no frontier model provider can deploy safely. Every model will face similar risks. The chilling effect is obvious: companies will either hide vulnerabilities (security through obscurity) or stop deploying advanced models to the public. Neither outcome serves actual security.
The hidden detail here is jurisdictional overreach. The US government is using export controls—designed for transistors and encryption—to control access to software services delivered over the internet. This works for citizens and companies based in the US, but it sets a precedent that will ripple globally. If the US can unilaterally disable Anthropic's models for "foreign nationals," the EU can require different restrictions, China can require model localization, and the internet fractures further. That outcome weakens rather than strengthens security.
The timing also matters. Three days after free deployment to millions of users, mid-rollout, is the worst possible moment to enforce this order operationally. It suggests either that the government's concern evolved after the rollout, or that enforcement lags detection. Either way, it raises questions about whether this was a measured decision or an emergency response to political pressure.
The real test will be what comes next. If Anthropic restores access within weeks (as it promises), this becomes a cautionary episode. If access remains suspended, we're watching the beginning of a much larger regulatory framework that could reshape how frontier AI models are deployed globally.
— HackWire Editorial
---
## Related Coverage