๐Ÿ•ท๏ธ
SetupAI
๐Ÿ‡ฌ๐Ÿ‡งStalecollected in 25m

Anthropic disables advanced AI models due to US export controls

Anthropic disables advanced AI models due to US export controls
PostLinkedIn
๐Ÿ‡ฌ๐Ÿ‡งRead original on The Guardian Technology
#export-control#national-security#regulationfable-5-and-mythos-5anthropicfable 5mythos 5

๐Ÿ’กUS government forces Anthropic to restrict model access; a major shift in AI export control and national security policy

โšก 30-Second TL;DR

What Changed

Anthropic must disable Fable 5 and Mythos 5 for all foreign nationals.

Why It Matters

This move signals a tightening regulatory environment for frontier AI models, potentially impacting global research collaboration and enterprise deployment for international teams.

What To Do Next

Review your team's access protocols for frontier models to ensure compliance with emerging US export control regulations.

Who should care:Enterprise & Security Teams

Key Points

  • โ€ขAnthropic must disable Fable 5 and Mythos 5 for all foreign nationals.
  • โ€ขThe US government issued an export control directive citing national security risks.
  • โ€ขConcerns center on the potential for the models to be used to discover software vulnerabilities.
  • โ€ขAnthropic was not provided with specific details regarding the nature of the security threat.

๐Ÿง  Deep Insight

Web-grounded analysis with 30 cited sources.

๐Ÿ”‘ Enhanced Key Takeaways

  • โ€ขThe US government's directive mandates the suspension of access to Fable 5 and Mythos 5 for all foreign nationals, both within and outside the United States, including Anthropic's own foreign national employees, leading Anthropic to disable the models for all customers to ensure compliance.
  • โ€ขAnthropic understands the government's concern to be related to a potential 'jailbreak' method for Fable 5, which could identify minor, previously known software vulnerabilities; however, Anthropic disputes the severity of this issue, noting that similar capabilities are present in other publicly available models.
  • โ€ขThe models, particularly Mythos 5, have demonstrated advanced cybersecurity capabilities during pre-release testing, autonomously discovering and exploiting thousands of previously unknown zero-day vulnerabilities across major operating systems and web browsers.
  • โ€ขThis export control directive is part of a broader and evolving US government strategy to regulate advanced AI, which has included previous and ongoing discussions about controlling AI chips, cloud-based access to AI services, and preventing the diversion of such technologies to 'countries of concern' like China.
  • โ€ขFable 5 and Mythos 5 are based on the same underlying AI architecture; the key distinction is that Fable 5, the public version, incorporates additional safeguards that redirect sensitive cybersecurity and biology-related queries to a less capable model (Claude Opus 4.8), whereas Mythos 5 has these safeguards lifted for a select group of vetted partners in programs such as Project Glasswing.

๐Ÿ› ๏ธ Technical Deep Dive

  • Fable 5 and Mythos 5 share the same core underlying architecture and intelligence, with the primary difference being the implementation of safety safeguards.
  • Fable 5 includes classifiers that automatically route sensitive queries related to cybersecurity and biology to a less powerful model, Claude Opus 4.8, to mitigate risks.
  • Mythos 5, designed for trusted partners, operates with these specific safeguards lifted in certain domains, allowing its full capabilities to be utilized.
  • Earlier Mythos-class models, such as Mythos Preview, demonstrated the ability to autonomously find and exploit zero-day vulnerabilities across major operating systems and web browsers.
  • The testing methodology for Mythos Preview involved an agentic scaffold where the model operated in an isolated container, reading code, hypothesizing vulnerabilities, running projects for confirmation, and generating bug reports with proof-of-concept exploits.
  • Mythos 5 achieves a score of 80.3% on SWE-bench Pro, an improvement over Mythos Preview's 77.8% and Opus 4.8's 69.2%.
  • On the 'Humanity's Last Exam' benchmark with tools, Mythos 5 scored 64.5%, surpassing Opus 4.8's 57.9%.
  • Both Fable 5 and Mythos 5 support a large context window of 1 million tokens.

๐Ÿ”ฎ Future ImplicationsAI analysis grounded in cited sources

The incident will likely accelerate the implementation of stricter identity verification and access controls for advanced AI models across the industry.
The government's directive to disable access for 'foreign nationals' and Anthropic's subsequent decision to disable for all customers due to the inability to differentiate users in real-time suggests that AI providers will need more robust mechanisms to verify user nationality and control access to comply with future regulations.
This event sets a precedent for direct government intervention in the commercial deployment of frontier AI models, potentially leading to more frequent regulatory actions based on perceived national security risks.
The government's direct order to pull a commercial model from the market, rather than the company's own choice, indicates a shift towards treating frontier AI as national security assets, which could lead to similar actions for other powerful models.
The scope of US export controls will likely broaden further to explicitly include advanced AI models and cloud-based access to these models, complicating global AI development and deployment.
The directive specifically targets access to AI models, building on previous discussions and attempts to control cloud-based AI services, indicating a broadening scope of export control beyond just physical semiconductors.

โณ Timeline

2022-10
US implements initial restrictions on semiconductor exports to China.
2023-07
US Department of Commerce highlights existing 'catch-all' export controls applicable to AI systems with military or WMD applications.
2024
US export controls on semiconductors are expanded.
2025-01
Biden administration issues the 'AI Diffusion Rule' for advanced AI chip exports, later rescinded by the Trump administration.
2025-05
US Department of Commerce announces heightened global due diligence for AI semiconductors and rescinds the Biden-era AI Diffusion Rule.
2026-04-07
Anthropic announces Project Glasswing, an initiative to use its Claude Mythos Preview model for defensive cybersecurity with partners.
2026-06-09
Anthropic launches Claude Fable 5 (public) and Claude Mythos 5 (restricted), both sharing the same core architecture but with different safeguard implementations.
2026-06-13
US government issues an export control directive, ordering Anthropic to suspend Fable 5 and Mythos 5 access for foreign nationals.
๐Ÿ‡ฌ๐Ÿ‡งRead original article on The Guardian Technology
๐Ÿ“ฐ

Weekly AI Recap

Read this week's curated digest of top AI events โ†’

๐Ÿ‘‰Related Updates

Same topic

Explore #export-control

Same product

More on fable-5-and-mythos-5

Same source

Latest from The Guardian Technology

AI-curated news aggregator. All content rights belong to original publishers.
Original source: The Guardian Technology โ†—