What happened with Anthropic's Claude Fable 5 AI model?

AI researcher "Pliny the Liberator" claims to have bypassed the safety guardrails built into Anthropic's Claude Fable 5 within 48 hours of its launch. The claim has sparked debate about whether advanced AI safety measures can reliably prevent misuse.

Why is Claude Fable 5 different from Anthropic's Mythos model?

Fable 5 was released as a heavily restricted version of Anthropic's more powerful Mythos model. Anthropic previously said Mythos was too capable and potentially risky for broad public access without additional safety controls.

What is an AI jailbreak?

An AI jailbreak is a technique used to bypass a model's built-in restrictions and safety systems. Users attempt to make the AI respond to requests it would normally refuse, often through creative prompting methods.

What techniques did Pliny claim were used to bypass Fable 5?

According to Pliny, the jailbreak relied on methods including Unicode manipulation, homoglyph substitutions, narrative framing, long-context prompts and decomposition-recomposition techniques. These approaches were designed to make restricted requests appear harmless to safety filters.

What is decomposition-recomposition in AI prompting?

Decomposition-recomposition involves breaking a complex request into multiple smaller prompts that appear benign on their own. The outputs are then combined to achieve a result that may have been blocked if requested directly.

Why are crypto users paying attention to this story?

Some crypto users have previously expressed concerns that advanced AI systems could help identify software vulnerabilities or exploit weaknesses in blockchain protocols. A successful jailbreak could increase concerns about AI-assisted cyberattacks against crypto projects.

Has Anthropic confirmed the jailbreak claim?

No. At the time of publication, Anthropic had not publicly responded to Pliny's specific claims. The company previously stated that extensive internal testing and an external bug bounty programme found no universal jailbreak methods.

Image for illustrative purposes only. Not a real photo.

AI researcher claims Anthropic Fable 5 jailbreak

Written by Liezl GambePublished Jun 11, 2026, 6:44 AM

An artificial intelligence researcher known as "Pliny the Liberator" claims to have bypassed the safety guardrails of Anthropic's newly released Claude Fable 5 model less than two days after its launch.

Fable 5 was introduced on Tuesday as a heavily restricted version of Anthropic's more capable Mythos model, which the company previously described as too powerful and potentially dangerous for broad public release.

“Despite this overly sensitive, authoritarian ‘safety’ layer on top of Mythos, my lil liberators have been hard at work.. cleverly finding the holes in the fence that the thought police missed,”

Said Pliny.

The researcher said the jailbreak relied on techniques including Unicode and homoglyph manipulation, long-context prompting, narrative framing and decomposition-recomposition methods that break complex requests into smaller, seemingly harmless prompts.

According to Pliny, decomposition and recomposition proved particularly effective because individual prompts appeared benign to safety systems but could be combined to generate restricted outputs when assembled together.

The claims have intensified criticism of Fable 5's restrictive design, which redirects users seeking information on sensitive subjects such as cybersecurity or bioweapons to an older and less capable AI model.

“This is one of the first times that an AI company has rolled out a guardrail, and there has been uniform disdain,”

Said Princeton University AI researcher Sayash Kapoor, according to the Wall Street Journal.

Anthropic said before launch that an external bug bounty programme and more than 1,000 hours of testing failed to uncover any universal jailbreak methods, while the company had not publicly responded to Pliny's claims at the time of publication.