U.S. Imposes Export Restrictions on Anthropic AI Models Amid 'Jailbreak' Dispute

The U.S. Commerce Department has imposed strict export restrictions on Anthropic's new AI models, Fable 5 and Mythos 5. These controls, enacted late Friday, bar the distribution and use of the models by individuals outside the U.S. and foreign nationals within the U.S., including Anthropic’s non-citizen employees. Anthropic stated it had no alternative but to disable the models for all users due to the directive's broad scope.

Anthropic indicated that the Commerce Department's action was spurred by research into a “jailbreak” of its AI, which sought to probe the bypassing of safeguards. The company expressed disagreement with this assessment, arguing that a “narrow potential jailbreak” should not warrant recalling a commercial model widely deployed. Anthropic asserted that applying such a standard across the industry could effectively halt all new model deployments for frontier AI providers. The company acknowledged the government’s right to block unsafe AI but contended that such actions should follow a transparent, fair, and technically sound statutory process, which it believes this action did not.

Katie Moussouris, CEO of cybersecurity firm Luta Security, viewed the research findings, which were reportedly produced by Amazon researchers using prompts to identify security vulnerabilities. Moussouris stated that the research was “not a jailbreak” but rather “Defense Oriented Prompting (DOP),” which she described as capabilities essential for defenders. She further commented that if national defense is the objective, this action represents “an own goal” for the U.S.

The administration’s directive, specifically prohibiting foreign nationals in the U.S. from using Anthropic’s new models, has raised concerns among experts. Ben Murphy, a scholar at the Institute for Progress think tank, characterized the move as “another step on the balkanization of technology.” Murphy noted that while requiring proof-of-citizenship for service access was once uncommon, it is becoming increasingly frequent with new technologies. He also highlighted the unpredictability of such actions, suggesting that AI labs might be prompted to keep models in-house or become less transparent with the government about potential vulnerabilities, citing Anthropic's transparency seemingly backfiring.

According to Fortune, Anthropic had previously been identified by the administration as a supply-chain risk for Pentagon contractors, despite providing early access to its Mythos model and warning about its cybersecurity implications.