Anthropic announced it will immediately halt user access to its cutting-edge AI models following a US government directive that restricts these models from being accessed by foreign nationals, citing national security issues.
The company revealed it was informed of the export control measure, which mandates the suspension of its Fable 5 and Mythos 5 models for foreign users, although specific national security details were not disclosed. Anthropic noted that the government’s concerns stem from a belief that there exists a workaround or “jailbreak” that could enable Fable 5 to be exploited for uncovering software vulnerabilities.
This order comes amidst a thaw in the previously contentious relationship between Anthropic and the Trump administration during its efforts to prepare for a public offering. This year, tensions escalated after Anthropic declined to allow the US military to utilize its AI technologies for domestic surveillance and autonomous weaponry. In retaliation, the government blacklisted Anthropic from the supply chain, an action set to be enforced later this year.
The recent directive signifies a substantial uptick in US initiatives aimed at curtailing the AI capabilities of foreign adversaries. Historically, US export controls have concentrated on the hardware and tools powering AI rather than directly restricting foreign access to AI technologies.
Anthropic indicated that the only feedback they received from the government was verbal, referencing a possible, limited jailbreak. The company expressed its disagreement with the notion that the identification of this narrow potential jailbreak warranted the withdrawal of a model that is safely used by hundreds of millions.
The situation accentuates the growing friction between AI developers and regulatory bodies, particularly concerning how to evaluate risks associated with methods that circumvent the safeguards of these models. Just days prior, Anthropic had advocated for increased oversight in the AI sector, proposing measures to prevent models with unacceptable risks from being deployed. Nonetheless, they feel the recent government action did not adhere to fair and evidence-based regulation practices.
Kirsten Davies, the Pentagon’s chief information officer, remarked on X that national security should take precedence over financial gains. “Some things are simply more important than revenue cycles, clickbait and pre-IPO valuation. America First. Always,” she stated.
Anthropic is currently in the process of confidentially pursuing an IPO in the US, moving ahead of competitors like OpenAI in the race for public market entry. This week also saw the launch of Anthropic’s AI model Claude Fable 5, which boasts a new level of capability termed “Mythos-class,” although it features restrictions against usage in high-risk areas such as cybersecurity—limitations that some users have described as overly broad.
Experts warn that Mythos models could potentially accelerate advanced cyber threats if misused, especially in industries like banking that depend on intricate and often legacy technology networks. Anthropic claims to have collaborated with the US government and others to ensure safety prior to the Fable model's rollout and highlighted that models from other AI competitors exhibit similar capabilities in identifying minor coding issues.
To comply with the directive, Anthropic stated it must quickly deactivate Fable 5 and Mythos 5 for all users, although access to its other models remains unaffected. The company expressed hope for a resolution to what it perceives as a misunderstanding, aiming to restore access to its models in the near future. A US official confirmed that the commerce department had formally issued the directive to restrict all foreign access to Fable 5 and Mythos 5.
