Why the U.S. Lifted Curbs on Anthropic’s Fable 5 and Mythos 5

The abrupt shutdown—and fast reinstatement—of Anthropic’s Fable 5 and Mythos 5 showed how frontier large language models will now be governed: via rapid, high‑stakes negotiations between labs and Washington in the United States. [1][8]

For AI buyers, policy teams, and technical leads, the episode showed that model access can vanish overnight due to export controls, AI safety regulation, and AI security concerns—not just bugs or outages. [8][10] A new bargain is emerging: deeper safeguards, monitoring, and government visibility into high‑risk capabilities like cyber offense, in exchange for regulators keeping models online. [2][4]

• Key takeaway: Access to top‑tier LLMs now turns on real‑time national security judgments in a global AI race. [1][7]

From Sudden Suspension to Partial, Then Full, Reinstatement

On June 13, the Trump administration ordered Anthropic to block all foreign nationals from Fable 5 and Mythos 5, including non‑U.S. staff in the U.S. [8][10] Because Anthropic could not segment users by nationality at inference time, it shut down both Claude Fable 5 and Mythos 5 globally. [8]

Officials feared a newly publicized jailbreak could let hostile actors turn Fable 5 into an automated “vuln‑finder” for exploiting software, servers, and GPU systems at scale—potentially deployable by autonomous AI agents. [7][8]

Anthropic countered that:

• The jailbreak exposed only “a small number of previously known, minor vulnerabilities”

• Other public models could find similar issues without bypassing safeguards

• Fable 5’s cyber‑related outputs were already heavily restricted and extensively red‑teamed with government, allies, and third parties [8]

• Data point: The suspension lasted about two and a half weeks before Commerce eased controls. [1][3]

During that time, Anthropic negotiated with the Commerce Department. An initial compromise:

• Restored Mythos 5 to about 100 “trusted” U.S. organizations—Fortune 500 firms, critical infrastructure operators, cybersecurity partners, and federal agencies—under tight conditions [4][5]

• Kept Fable 5 offline pending further review [4]

• Key point: Export rules targeting foreign nationals effectively shut down models for everyone, revealing how blunt current compliance tools are in a fragmented global regime. [8][10]

What Changed: Safeguards, Oversight, and Government Confidence

Commerce Secretary Howard Lutnick said his department “worked closely” with Anthropic for roughly two weeks to analyze and approve Fable 5 and Mythos 5. [1][2] In a letter, Lutnick said Anthropic committed to:

• Proactively detect and mitigate security risks tied to both models
• Inform the government when it detects malicious activity involving them [2]

This:

• Creates ongoing monitoring and reporting channels for high‑risk behavior
• Extends pre‑launch red‑teaming into continuing operational oversight [2][4]

The Commerce Department lifted export controls on both models less than three weeks after the suspension, reflecting both intensive security review and pressure to restore advanced AI tools for the U.S. economy and its competitors. [1][9] Anthropic began restoring global access the next day. [1][3]

Lutnick’s letter preserved the right to reimpose curbs if risks change or Anthropic falls short. [2] This is conditional approval, not permanent clearance—and explicitly reversible. [2][9]

• Key point: Frontier model access is now subject to “regulatory rollback” risk, alongside technical rollback risk. [2][9]

Implications for AI Policy, Industry, and Global Users

The Fable/Mythos case fits a broader pattern: the Trump administration is actively vetting frontier releases, including asking OpenAI to stagger GPT‑5.6 Sol’s rollout to a short list of government‑approved partners before wider access. [6][7][9] This builds on Executive Orders like the January 2025 AI order and Ensuring a National Policy Framework for Artificial Intelligence, which formalized federal vetting for frontier Artificial Intelligence. [7][9]

Both Anthropic and OpenAI—whose ChatGPT, GPT‑5, and newer models power copilots and autonomous AI agents—have accepted limited‑partner previews shaped by Washington’s risk assessments. [6][9]

Earlier restrictions on Fable 5 and Mythos 5:

• Disrupted global AI R&D by sidelining foreign‑national staff from their own models [10]

• Forced multinationals to reconfigure access by citizenship across allied countries and China [4][10]

• Drove rapid tool migration for vulnerability discovery, code analysis, cyber defense, and AI in Education tools used by teachers and downstream adopters [4][10]

• Key takeaway: “Who can use which model?” is now an exports and export‑control question, not just pricing or tiering. [7][10]

Washington’s use of emergency export controls, limited‑access approvals, and an executive order enabling up to 30 days of national security review for frontier models signals a more assertive AI governance phase. [7][9] In a world defined by competition over semiconductors, servers and GPU systems, and AI exports with China and other competitors, frontier LLMs are treated as strategic infrastructure. Labs must bake regulatory timelines, reporting, usage limits, and evolving AI safety regulation into product and risk planning. [4][7]

For businesses, this raises questions:

• How might export controls on models like Fable 5 shift in the next year?

• Will conditional approvals and “trusted partner only” launches become standard?

• How should procurement, security, and policy teams hedge against sudden access changes?

• Key point: If core workflows depend on a single frontier model, you now carry regulatory concentration risk, not just vendor lock‑in. [6][9]

Conclusion: Conditional Reprieve, Not Regulatory Peace

Anthropic’s path from global shutdown to full reinstatement of Fable 5 and Mythos 5 compressed a new regulatory playbook into under three weeks: emergency export controls, joint security reviews, stronger safeguards, and formal commitments to ongoing monitoring and government notification. [1][2][8]

Curbs are lifted, but approval is conditional, politically sensitive, and reversible as capabilities and threats evolve—especially amid a fast U.S.–China AI race and shifting global rules. [2][7]

AI buyers, policy professionals, and technical leaders should:

• Track emerging U.S. rules on frontier models
• Reassess exposure to export‑controlled systems
• Diversify critical dependencies and plan for rapid compliance pivots