US Government Forces Anthropic to Block Foreign Access to Fable 5 and Mythos 5 — Even Its Own Non-Citizen Employees

In what appears to be the first instance of an AI company being ordered to restrict access to frontier models based on nationality (including inside the United States), the U.S. government has directed Anthropic to suspend access to its latest models, Fable 5 and Mythos 5, for any foreign nationals.

The directive, issued under national security and export control authorities, arrived late on June 12, 2026 (around 5:21 p.m. ET). Because Anthropic lacks real-time infrastructure to verify citizenship or nationality on the fly, the company took the drastic step of disabling both models for all users worldwide to ensure compliance. Other Anthropic models, including the widely used Opus series, remain available.

---

The Immediate Impact and Irony

The ban applies to foreign nationals both outside and inside the U.S., explicitly including Anthropic’s own employees who are not U.S. citizens.

This has created an unusual situation where researchers and engineers working directly on these models can no longer use them.

One high-profile example is Andrej Karpathy, the renowned AI researcher, OpenAI co-founder, and former Tesla Autopilot AI lead.

Karpathy joined Anthropic’s pre-training team in May 2026. Born in Slovakia and holding Canadian citizenship, he now finds himself unable to access the very models he is helping to develop and improve.

---

What Triggered the Order?

According to Anthropic, the government acted after learning of a method to bypass or “jailbreak” Fable 5’s safety protections. The company reviewed a demonstration of the technique and described it as narrow and non-universal: essentially, prompting the model to analyze a specific codebase and identify/fix vulnerabilities within it.

Anthropic noted that the jailbreak primarily surfaces a couple of minor, already-known bugs — issues that other publicly available frontier models (including OpenAI’s GPT-5.5) can also identify without any special jailbreak.

The company emphasized that its extensive red-teaming (thousands of hours, including with U.S. and U.K. government partners) and layered safeguards make the risks comparable to those of existing models.

---

Fable vs. Mythos: Same Core, Different Guardrails

Fable 5 and Mythos 5 are based on the same underlying model.

The key difference lies in the safety layers:

• Fable 5 is the public-facing version with stricter guardrails (particularly around cybersecurity, biology, chemistry, and model distillation) to make advanced capabilities safer for broader release.
• Mythos 5 has lighter or fewer restrictions and is available only to vetted organizations and partners (e.g., through programs like Project Glasswing for cybersecurity work).

Users have previously complained that Fable’s guardrails can be overly sensitive — mentioning cybersecurity topics sometimes triggers refusals even for legitimate defensive research.

---

Anthropic’s Response and Broader Context

Anthropic has called the situation a “misunderstanding” and stated it is working urgently to restore access. The company argues that pulling a widely used frontier model over one narrow, limited technique is disproportionate and, if applied industry-wide, would severely hinder the deployment of advanced AI systems.

This move fits into a longer pattern of U.S. efforts to control the spread of advanced AI and computing technology. Earlier attempts, such as the Biden-era “AI Diffusion” framework (which aimed to tier countries and control access to model weights), were rescinded in 2025 as overly bureaucratic. The current action represents a more targeted, enforcement-heavy approach applied directly to a leading AI lab.

The parallel with hardware export controls is telling: non-U.S. companies have long faced significant hurdles acquiring the most advanced Nvidia GPUs. Extending similar logic to model access marks a significant escalation.

Also read:

• What Is Physical AI and How Jeff Bezos’ Startup Prometheus Aims to “Hack” the System
• AI Fuels Surge in Solo Entrepreneurs, Nasdaq Study Reveals
• The Great App Icon Divide: Clean Minimalism in the West vs. Dynamic Marketing in China
• Korea's Gen Z Fights Delivery Cravings with 'Dopamine' Dummy Apps

---

What Comes Next?

If this policy holds or expands, we could see frontier model access increasingly tied to nationality or citizenship verification — not just at Anthropic, but potentially at OpenAI, Google, xAI, and others. China and other nations may respond with their own restrictions on top-tier models, accelerating the move toward “sovereign” or regionally siloed AI systems.

A key practical question is implementation: simple IP changes or VPNs will likely become insufficient. More robust Know-Your-Customer (KYC) processes for AI access could emerge, though they risk creating friction for the vast majority of legitimate users while sophisticated actors (those seeking to distill models or misuse capabilities) find workarounds — much like how weapons restrictions fail to stop determined criminals.

Anthropic has apologized for the disruption to customers and remains optimistic about a relatively swift resolution through dialogue with authorities. Whether that optimism proves justified remains to be seen.

This episode highlights the growing tension between rapid AI advancement, commercial deployment, and national security imperatives. As frontier models become more powerful, governments are clearly willing to intervene more directly — and the ripple effects are already being felt inside the companies building them.

The coming weeks will show whether this is an isolated “misunderstanding” or the beginning of a new, more restrictive era for global access to cutting-edge AI.