Anthropic has removed two of its language models, Fable and Mythos, from public use after the U.S. government imposed export restrictions. The decision followed a report from Amazon that raised concerns about the models' vulnerability to jailbreak attempts. The company acted quickly to comply with the new regulations.
The export restrictions were triggered by the U.S. government’s assessment that the models could be manipulated to bypass safety controls. Amazon’s report highlighted specific behaviors that could enable users to override content filters. The government’s action aimed to limit the potential misuse of advanced AI systems.
Anthropic contested the details of Amazon’s findings, stating that it had already addressed the vulnerabilities. The firm introduced an enhanced safety classifier designed to detect and block the problematic behaviors cited in the report. The update was rolled out promptly to mitigate the identified risks.
Earlier, Fable had faced criticism for its handling of sensitive requests. The model frequently rejected queries related to cybersecurity and biology, and it redirected users away from certain topics without providing explanations. Users reported a lack of transparency regarding the reasons for the rejections.
In response, Anthropic announced that routine tasks such as coding and debugging would revert to its Opus 4.8 model, which offers lower computational power. The shift reflects a cautious approach to balance functionality with safety. Developers will need to adjust to the new model’s capabilities.
The removal of Fable and Mythos and the transition to Opus 4.8 underscore the growing scrutiny of AI safety features. The changes may affect developers who rely on advanced language models for complex tasks. The incident highlights the need for robust safety mechanisms and clear communication from AI providers.
