AI Researcher Claims Anthropic’s Claude Fable 5 Guardrails Were Bypassed Within Days

An artificial intelligence and cybersecurity researcher known as “Pliny the Liberator” claims to have bypassed the safety protections built into Anthropic’s newly launched Claude Fable 5 model less than 48 hours after its release.

Claude Fable 5 was introduced as a heavily safety-focused version of Anthropic’s more powerful Mythos model. The company limited access to Mythos and designed Fable 5 with stricter controls to prevent users from obtaining potentially harmful information, including hacking methods and dangerous scientific instructions.

Techniques Used to Bypass AI Guardrails

According to Pliny, several methods were used to work around the model’s restrictions. These included Unicode and homoglyph manipulation, long-context prompts, narrative framing, academic-style decomposition and recomposition, and assistance from a jailbroken version of Claude Opus 4.8.

Pliny said decomposition and recomposition was particularly effective. The technique breaks a sensitive request into multiple harmless-looking questions, which can later be combined to produce information that safety systems are designed to block.

Critics Question Fable 5 Restrictions

Since launch, Claude Fable 5 has faced criticism from some researchers who argue that the model’s restrictions limit legitimate research. When users ask about certain cybersecurity or scientific topics, the system redirects them to a less capable model instead of providing detailed responses.

Anthropic Says Testing Found No Universal Jailbreak

Before release, Anthropic conducted internal evaluations and an external bug bounty program to identify weaknesses. The company stated that more than 1,000 hours of testing failed to uncover any universal jailbreak capable of bypassing all safeguards.