Anthropic and Washington: The Vulnerability of LLMs Challenged

Anthropic Faces Criticism from the U.S. Government

The U.S. government has recently expressed concerns regarding Anthropic, a company specializing in the development of artificial intelligence models. Authorities accuse Anthropic of having ignored an executive order from the Trump administration by releasing the Fable 5 model without prior approval. This order, which called for government oversight of AI models, had initially been well-received by Anthropic. However, the company chose to launch Fable 5 before the designated approval channel was operational.

A government official summarized the conflict between the Trump administration and Anthropic by stating that Anthropic had misled them. "Everyone said that Anthropic was a malicious actor. Some of us said it was time to give them a chance. Now, those people are questioning that. They misled us."

Communication and Security Challenges

Discussions between the U.S. Department of Commerce and Anthropic are ongoing, but the exchanges are deemed complex, with both parties seemingly speaking different languages. Additional meetings are planned with the participation of the CIA and Michael Kratsios, the science advisor. This situation highlights a lack of understanding on the part of the government regarding the security of AI models.

The accusation that Anthropic was aware of the risks of "jailbreak" brings to light a significant challenge. Experts, including OpenAI, have already warned that hacking methods such as prompt injection may never be completely secure. Anthropic's CEO, Dario Amodei, stated in 2023 that a "jailbreak could be a matter of life or death" if someone managed to bypass security protocols in science, technology, and biology.

The real question is how serious the breach is and how quickly countermeasures are implemented. However, if the U.S. government insists that cutting-edge AI models be "invulnerable" before their international shipment, difficult discussions lie ahead.

Call for Lifting Export Restrictions

In response to the accusations, over 100 cybersecurity experts have signed an open letter addressed to Commerce Secretary Lutnick and National Cyber Director Cairncross. They are calling for the lifting of export controls on the Fable and Mythos models, arguing that these restrictions deprive cybersecurity defenders of valuable tools.

The signatories, including figures like Alex Stamos, Rachel Tobac, Katie Moussouris (Luta Security), Dan Lorenc (Chainguard), and Joe Levy (Sophos), emphasize that other models, such as GPT-5.5, Opus, Sonnet, and the Chinese Kimi 2.7, offer comparable performance. They believe that the security measures integrated into Fable are already considered excessive by the security community at the time of its launch.

The current situation highlights the tensions between the need to regulate cutting-edge technologies and the importance of not stifling innovation in the field of artificial intelligence. Experts warn that open-weight Chinese models are just months away from matching the best American models, which could have significant implications for the technological competitiveness of the United States.