Cybersecurity Experts Protest US Ban on Anthropic’s Fable and Mythos Models

A coalition of 76 leading cybersecurity veterans is sounding the alarm against a recent U.S. government export control order targeting Anthropic’s most advanced AI models. The group argues that by restricting access to these tools, the government is inadvertently disarming digital defenders while adversaries continue to advance.

The Conflict: National Security vs. Defensive Capability

The tension began when the U.S. government issued an export control order on Anthropic’s Fable and Mythos models, citing undisclosed national security concerns. In compliance with the order, Anthropic has suspended worldwide access to these models.

The Mythos model was originally designed with such high-level capabilities for vulnerability discovery that Anthropic initially limited access to a select group of roughly 150 organizations across 15 countries. The public-facing version, Fable, was intended to include strict guardrails to prevent misuse in biology, chemistry, and cybersecurity. However, the government's intervention has effectively halted the deployment of these powerful tools for the very people tasked with securing the internet.

The "Jailbreak" Controversy and the Amazon Paper

Anthropic suggests the White House order may stem from concerns regarding "jailbreaking"—methods used to bypass safety guardrails. This concern reportedly originates from a non-public research paper by Amazon researchers.

The paper suggested that users could bypass Fable’s security restrictions to access Mythos-level capabilities. However, cybersecurity experts, including Katie Moussouris (founder of Luta Security), argue this is a fundamental misunderstanding of AI utility. Moussouris contends that the "jailbreak" described was simply the model performing its intended function: fixing open-source code containing known vulnerabilities.

According to Moussouris, asking an AI to fix a bug, explain the patch, and write a test is not a security breach; it is the "find, fix, and test loop" that defines modern defensive security. Attempting to block these behaviors would fundamentally cripple the model's ability to protect software.

High-Stakes Signatories and Industry Implications

Otwarty list wspierają autorytety w społeczności bezpieczeństwa, w tym były szef ds. bezpieczeństwa w Facebooku Alex Stamos, założyciel Bugcrowd Casey Ellis oraz słynny kryptograf Jon Callas. Ich argumentacja koncentruje się na krytycznej nierównowadze: jeśli obrońcom odmówi się dostępu do najnowocześniejszych modeli LLM, podczas gdy przeciwnicy będą korzystać z nieograniczonych modeli, globalny poziom bezpieczeństwa ulegnie osłabieniu.

Eksperci zauważyli również, że postrzegana podatność w Fable nie jest unikalna dla Anthropic. W liście sugeruje się, że podobne „podatności” mogą wystąpić w GPT-5.5 od OpenAI, własnych modelach Anthropic, takich jak Claude Opus 4.8 i Sonnet, a nawet w modelach międzynarodowych, takich jak Kimi 2.7.

Grupa wzywa do demokratycznego, opartego na nauce procesu regulacyjnego, który opiera się na przejrzystych badaniach, a nie na szerokich, reaktywnych zakazach, które mogą przynieść więcej szkody niż pożytku.

Kluczowe wnioski