Claude Fable 5 returns after US export controls ease
Anthropic has brought back its most capable generally available model, Claude Fable 5, after the company confirmed that US export controls were lifted on July 1. The temporary suspension began on June 12 following a directive restricting access based on nationality verification challenges, which Anthropic could not perform in real time. Mythos 5, a related model with fewer safeguards for certain US cybersecurity partners, remains partially restricted.
A targeted fix for a reported bypass
The pause followed an Amazon report detailing a method to bypass Fable 5’s safeguards, prompting the model to identify and even demonstrate software vulnerabilities. In response, Anthropic developed a new safety classifier designed to block the specific technique in over 99% of cases. Requests flagged by the classifier are not rejected outright but rerouted to Claude Opus 4.8, with users notified of the fallback. The Department of Commerce’s CAISI team evaluated both old and new safeguards and concluded the updated protections are extraordinarily strong, though they acknowledge a tradeoff in higher false positives during routine coding tasks.
Context and industry implications
Anthropic emphasized that the bypass technique was not unique to Fable 5. Tests showed that other leading models—including Claude Opus 4.8, GPT-5.5, and Kimi K2.7—could replicate similar outputs under the same conditions. The incident has also highlighted the need for a standardized framework to assess jailbreak severity across the industry. Anthropic’s approach reflects a layered defense strategy, where smaller classifiers act as early detectors within a broader safety margin that may occasionally block benign requests.
For users, Fable 5 is now accessible globally on the Claude Platform, Claude.ai, Claude Code, and Claude Cowork, with Mythos 5 access restored to a limited set of US organizations. The episode underscores both the evolving nature of AI safety and the regulatory realities shaping global AI deployment.
Source: MarkTechPost. AI-assisted editorial synthesis — TechnoExpress.