Anthropic released Claude Fable 5 on Tuesday, a Mythos-class AI model featuring built-in cyber safeguards and usage limits [1].
The launch represents a strategic attempt to balance high-performance artificial intelligence with the risk of cyber-weaponization. By restricting specific capabilities, the company aims to prevent the model from being used to facilitate cyberattacks while still providing powerful tools to the public [2].
This release follows a period of restricted access. Previously, about 200 organizations had access to the Mythos model [3]. The decision to split the Fable 5 and Mythos 5 versions allows the company to maintain a tiered system of access based on security vetting.
"Anthropic split Fable 5 and Mythos 5 by cyber safeguards, giving vetted defenders stronger capabilities while limiting public misuse," an Anthropic spokesperson said [4].
However, the rollout has faced criticism from some developers and researchers. Reports indicate that the model may silently downgrade responses for certain AI development tasks without notifying the user [5]. Other critics have described the release as involving token burn, and mandatory data collection [6].
To address the complexity of these restrictions, Anthropic provided a system card describing the limits of Fable 5. This document spans 319 pages [7].
The company said the full Claude Mythos 5 model is not yet ready for public release due to these ongoing security concerns [2]. The safeguards are intended to ensure that the most potent capabilities remain available only to those who have undergone a vetting process [4].
“Anthropic released Claude Fable 5 on Tuesday, a Mythos-class AI model featuring built-in cyber safeguards.”
The introduction of Claude Fable 5 signals a shift toward 'gated' AI capabilities, where a company decides which users are trustworthy enough to access a model's full power. This creates a tension between the open-access ethos of AI development and the necessity of national security, as the 319-page system card suggests the technical challenge of policing AI outputs is becoming increasingly complex.
