Anthropic's "Fable 5" Resumes Global Distribution Following Government Ban

Anthropic's AI model "Fable 5" has resumed global distribution after approximately two weeks of suspension by the US government. The suspension was triggered by the discovery of a "jailbreak" technique that deliberately circumvents the model's safety restrictions. Anthropic has implemented new safety measures and obtained government approval to resume distribution.

A jailbreak refers to the act of extracting harmful content or dangerous information that an AI model is designed not to respond to, through specific input manipulation. This technique was discovered by Amazon researchers and was not limited to Fable 5. According to Anthropic's explanation, the smaller model "Claude Haiku 4.5" was also vulnerable to the same technique. In other words, the vulnerability was not unique to Fable 5 but rather a technical issue common to Anthropic's models as a whole.

To address this issue, Anthropic implemented a new "safety classifier." A safety classifier is a filtering function that determines whether input text has harmful intent. According to the company, this classifier can block the problematic technique in over 99% of cases. However, the company acknowledges that it also causes "false positives," mistakenly restricting some benign user requests.

The situation illustrates the difficult balance between releasing cutting-edge AI models and maintaining safety. The US government's decision to temporarily suspend distribution of an AI model demonstrates that oversight of AI safety is actually functioning. On the other hand, the fact that similar vulnerabilities existed in smaller models suggests that stopping a specific model alone may not provide a fundamental solution.

AI safety measures are still in development, and improving filtering accuracy while maintaining usability often involves a trade-off. The problem of blocking benign requests when enhancing safety, as seen in this case, is a common challenge across the industry. How Anthropic and other companies continue to balance these concerns remains an important point to watch.

The focus going forward will be on how much Anthropic's new safety classifier can improve in accuracy and reduce false positives. Additionally, the institutional question of how much government agencies should be involved in safety reviews of AI models may also influence future industry trends.