Anthropic’s latest model, Claude Fable 5.1, has cracked a cipher that remained unsolved for nearly four centuries. In 44 minutes and entirely without human guidance, the AI decoded a 64-number puzzle left by Scottish nobleman Thomas Urquhart in 1653—uncovering a hidden eight-line loyalty poem. This isn’t just a historical trivia win; it’s a powerful demonstration of how modern agentic AI can autonomously reason through complex, multi-layered problems.

What Happened

The cipher, known as the Cyphral Distich, was appended to Urquhart’s 1653 work Logopandecteision. It consists of two lines, each containing 32 numbers. For 373 years, cryptographers considered it an open mystery—formally listed by the journal Notes and Queries in 1899. The key insight: the numbers referenced 32 specific paragraphs in a source book, which served as the decryption key.

AI evaluation firm Vals AI tasked Claude Fable 5.1 with the puzzle. The model worked unsupervised—it had to infer the referencing system, locate the correct source text, map each number to its paragraph and word, and reconstruct the hidden message. The result was a loyalty poem to Charles II. Within the same session, the model also decoded a separate, related cipher from Urquhart’s 1652 work.

Read the full announcement →

My Take

This is a genuine milestone for agentic AI. Unlike benchmarks that test narrow skills (like code generation or trivia), this challenge required long-term planning, iterative hypothesis testing, and cultural-historical context—things models have traditionally struggled with. The fact that Claude did it in under an hour, unsupervised, signals a real leap in autonomous reasoning capability.

For developers and security researchers, the implication is clear: AI can now handle cryptanalysis that was previously the domain of human experts and months of effort. But it also raises questions. If a model can crack a 300-year-old cipher with minimal input, what does that mean for modern encryption and data privacy? The same reasoning could be applied to contemporary obfuscation, security audits, or even reverse engineering. This is both a powerful tool and a new attack surface to consider.

What to Watch

  • Historical archives: expect a wave of AI-driven decryption of long-lost manuscripts and codes.
  • Security tooling: agentic AI will likely become a standard part of penetration testing and vulnerability discovery.
  • Model benchmarking: “cryptographic reasoning” may become a standard metric alongside math and coding.