Anthropic pauses new AI model releases amid government national‑security warnings

0
62

Key Takeaways

  • The U.S. Department of Commerce ordered Anthropic to suspend foreign access to its two newest AI models, Fable 5 and Mythos 5, citing national‑security concerns over a reported jailbreak technique.
  • The export‑control directive bars use by any foreign national, including Anthropic’s own foreign‑born employees, prompting the company to disable the models entirely while leaving other models unaffected.
  • Anthropic contends the alleged vulnerability is narrow, non‑universal, and already present in publicly available models such as OpenAI’s GPT‑5.5, arguing the government’s reaction is disproportionate.
  • The move is part of an ongoing tension between Anthropic and the Trump administration, which has previously sought to limit the company’s federal contracts while simultaneously granting the NSA access to Mythos 5 for offensive cyber work.
  • Industry observers warn the decision could set a precedent for broad government intervention in AI model releases, affecting innovation, cybersecurity practices, and international collaboration.

Overview of the Order
On Friday afternoon the U.S. Secretary of Commerce, Howard Lutnick, issued a letter to Anthropic CEO Dario Amodei directing the company to immediately halt foreign access to its two most advanced models, Fable 5 and Mythos 5. The letter cited national‑security concerns tied to a method that purportedly bypasses the models’ built‑in safety guardrails. Because the restriction applies to any foreign national—whether located inside or outside the United States—Anthropic’s own foreign‑born engineers are also barred from using the systems. The directive took effect immediately, and Anthropic announced later that evening it had disabled the two models to ensure compliance, while leaving access to its other AI offerings unchanged.

Details of the Export‑Control Directive
The Commerce Department’s action placed Fable 5 and Mythos 5 under export controls that prohibit their use, transfer, or provision to any foreign person. This includes not only overseas users but also domestic employees who are not U.S. citizens or permanent residents. The letter did not disclose the exact nature of the security threat, instead relying on verbal evidence provided to Anthropic. By invoking the Export Administration Regulations, the government signaled its willingness to treat certain AI capabilities as strategic commodities akin to dual‑use technologies, a move that could reshape how model providers manage international distribution.

Anthropic’s Immediate Response
In a blog post released Friday night, Anthropic said it had reviewed the basis of the government’s concern and found the demonstrated capability to be a “narrow, non‑universal jailbreak.” The company asserted that the technique—prompting the model to read a specific codebase to identify software flaws—does not enable broad circumvention of safety measures and is already replicable using other publicly available models. Consequently, Anthropic disagreed with the severity of the finding, arguing that recalling a commercial model deployed to hundreds of millions of users over such a limited issue would set an untenable precedent for the entire frontier‑model industry.

Description of Fable 5, Mythos 5, and Project Glasswing
Fable 5 and Mythos 5 were unveiled earlier in the week as Anthropic’s most capable systems to date, representing the forefront of its research into large‑scale language models. Mythos 5 was made available to participants in Project Glasswing, a program that lets select cybersecurity firms employ the model to detect and remediate vulnerabilities in software. The precise impact of the Commerce Department’s order on Project Glasswing remains unclear, as Anthropic did not respond to requests for comment on whether the program’s access was revoked or merely restricted to U.S.‑only participants.

The Alleged Jailbreak Technique and Anthropic’s Dispute
According to Anthropic, the government’s evidence centers on a method termed “Defense Oriented Prompting,” whereby natural‑language instructions are crafted to treat the model as if it were executing code, thereby coaxing it to reveal software defects. Katie Moussouris of Luta Security highlighted this approach on BlueSky, noting it reflects a security‑first mindset that treats prompts as code. Anthropic maintained that the jailbreak is narrow—affecting only a specific class of prompts—and that similar capabilities are routinely used by cybersecurity professionals for legitimate defensive testing, including with models like OpenAI’s GPT‑5.5. The company emphasized that no tester has discovered a universal jailbreak capable of broadly disabling the models’ safeguards.

Expert and Industry Reaction
The decision drew skepticism from several analysts. Dean Ball, a senior fellow at the Foundation for American Innovation, labeled the move “baffling,” questioning the proportionality of restricting a model over a limited vulnerability. Chris McGuire of the Council on Foreign Relations acknowledged that targeted export controls could be a legitimate policy tool but condemned the blanket restriction as “highly questionable,” especially the deemed‑export provisions that affect foreign nationals working within the United States. Conversely, some government officials, including DOD CIO Kirsten Davies, voiced support, framing the action as a necessary prioritization of national security over commercial interests.

Political Context with the Trump Administration
The order continues a pattern of friction between Anthropic and the Trump administration. In February, President Donald Trump sought to bar Anthropic’s products from federal agencies after the company advocated for stricter limits on Pentagon use of its technology. Despite that, the NSA was later granted access to Mythos 5 under Project Glasswing for offensive cyber operations. Earlier this month, Trump signed an executive order urging federal agencies to strengthen cyber defenses and to create a voluntary mechanism for early government access to powerful AI models before public release, underscoring the administration’s dual focus on restricting and leveraging advanced AI.

Implications for AI Regulation and the Industry
Aaron Levie of Box characterized the directive as “a big turning point for AI regulation,” warning that treating specific models as too dangerous for certain uses could establish a precedent allowing the government to halt or reshape future model releases across the sector. Such a shift could impede innovation, increase compliance burdens for AI developers, and complicate international collaborations that rely on shared access to cutting‑edge models. Industry leaders cautioned that overbroad restrictions might push research offshore or incentivize the development of alternative, less‑transparent systems that evade regulatory scrutiny.

Outlook and Anthropic’s Position
Anthropic stated it believes the situation stems from a misunderstanding and is working with officials to restore access as soon as feasible. The company reiterated its commitment to a “defense‑in‑depth” strategy for model safety, combining narrow jailbreak resistance with continuous monitoring. While the immediate effect is the suspension of Fable 5 and Mythos 5 for foreign users, the longer‑term consequences will hinge on how the administration balances national‑security imperatives with the need to sustain a vibrant, globally competitive AI ecosystem. The episode serves as a litmus test for how governments may intervene in the rapidly evolving landscape of frontier artificial intelligence.

SignUpSignUp form

LEAVE A REPLY

Please enter your comment!
Please enter your name here