US Blocks Foreign Access to Anthropic’s Fable 5 and Mythos 5
US export controls disable Anthropic’s top models globally over a jailbreak scare—builders face sudden compliance costs.
The story of the day
The US government just forced Anthropic to disable Fable 5 and Mythos 5 worldwide after an export control directive cited national security risks tied to a potential safeguard bypass. Anthropic complied but disputes the rationale: the demonstrated jailbreak was minor, already known, and replicable with other public models like GPT-5.5. The move strands developers using these models, while raising the question of whether recalling commercial AI over narrow vulnerabilities sets a precedent for freezing innovation. For builders, the immediate cost is operational—retooling workflows around suddenly unavailable models. The longer-term risk: a new class of compliance overhead for frontier model access.
New tools & launches
Validates AI — Open-source tool for validating AI outputs against custom rules. [validates.ai]
Whissle Gateway — 500MB Docker container for local multi-modal voice AI inference. [whissle.ai/gateway]
Contrail Compute AIX — First RISC-V-based AI execution platform for edge deployment. [epicsemi.com]
Models & research
Anthropic’s Fable 5 and Mythos 5 are now inaccessible globally after a US export control order, despite the company’s claim that the cited jailbreak was non-critical. The models, used by hundreds of millions, were disabled for all users—foreign and domestic—to comply with the directive. Proprietary, now offline.
Industry pulse
Anthropic publicly pushed back on the US government’s decision, arguing that disabling commercial models over a known, minor vulnerability could "effectively freeze new model development." The company’s statement underscores the tension between security mandates and the practical costs of abrupt deprecation for developers.
Stat of the day
A German court ruled Google liable for false statements generated by AI Overviews, setting a precedent that platform operators assume legal responsibility for AI outputs. [Wired]
Get the full analysis — subscribe to Gradient Pro.
