US blocks Anthropic's powerful new AI models over security fears

The U.S. government has issued an export control directive, ordering AI firm Anthropic to immediately block access to its latest flagship models, Claude Fable 5 and Claude Mythos 5. This decision, driven by national security concerns under the Trump administration, has sent ripples through Silicon Valley, casting doubt on the future of commercial generative AI development.

Anthropic was compelled to completely disable both systems for all clients due to their global accessibility. The models, launched just three days prior on June 9, were hailed as technological breakthroughs. Fable 5, positioned as a competitor to OpenAI's GPT-5.5 and Google's Gemini, demonstrated unprecedented capabilities in programming and complex autonomous tasks.

The Mythos architecture, however, generated significant controversy even before its release. A preview version proved so effective at automatically detecting and exploiting system vulnerabilities during April lab tests that Anthropic itself described it as "too powerful to release." Despite safety filters implemented in the commercial Fable 5 version, the Trump administration deemed the cyber risks too substantial.

While Washington has not provided detailed technical evidence, Anthropic's official statement indicates the Department of Commerce received a report demonstrating a method to bypass Fable 5's security. This method allegedly allowed the model to analyze codebases and identify critical vulnerabilities, potentially exploitable by malicious hackers targeting critical infrastructure like energy or banking systems.

Anthropic developers strongly contest these allegations, emphasizing that the system underwent extensive safety testing prior to its release. The company highlighted that the Fable 5 model, in particular, passed thousands of safety tests. The U.S. government's move raises questions about its strategy for maintaining a competitive edge in the global AI race, potentially reserving the most advanced AI capabilities for domestic use.

“

Departament Handlu wszedł w posiadanie raportu demonstrującego metodę obejścia zabezpieczeń (jailbreaking). Miała ona pozwalać modelowi Fable 5 na analizę baz kodu i wskazywanie krytycznych podatności, co potencjalnie mogłoby zostać wykorzystane przez wrogich hakerów do ataków na infrastrukturę krytyczną, np. energetyczną czy bankową.