Anthropic Unveils CryptanalysisBench as Claude Finds Crypto Flaws

Anthropic announced a major breakthrough by its Claude models in the field of cryptography. The AI can now identify mathematical flaws within cryptographic algorithms themselves, rather than merely spotting code implementation bugs. To evaluate the cryptanalysis capabilities of large language models, the company released the CryptanalysisBench benchmark, revealing that frontier models can break 65%–86% of Tier-1 cryptographic schemes.

Confirmed

In collaboration with ETH Zurich, Tel Aviv University, and the University of Haifa, Anthropic introduced CryptanalysisBench, a benchmark specifically designed to test the cryptanalysis capabilities of LLMs.
Benchmark tests indicate that frontier models are capable of breaking 65%–86% of Tier-1 cryptographic schemes.
Claude discovered a new attack method targeting the post-quantum digital signature scheme HAWK, significantly weakening its security.
Claude also identified new vulnerabilities in reduced-round AES.

Unconfirmed

The specific model name Claude Mythos (or Claude Mythos Preview) mentioned in the post primarily stems from Polymarket relays and community rumors. While treated as a significant signal of the model series' capabilities, the exact official model name released by Anthropic remains to be confirmed.

Why It Matters

As @moyix pointed out, AI has previously been used mostly to find implementation vulnerabilities in software code. However, this research demonstrates that AI can now autonomously discover underlying mathematical flaws in cryptographic algorithms, including attacks on post-quantum cryptography candidates.
This progress indicates that AI shows immense potential in advanced security and cryptography research, even uncovering weaknesses long overlooked by human experts. In the future, it could profoundly shift the paradigm of security offense and defense, as well as cryptographic research.

2026-07-29 ~ 2026-07-29 · 7 related posts

Primary sources

Anthropic says Claude found new weaknesses in HAWK and round-reduced AES — AnthropicAI · 2026-07-29 01:16 (UTC+08:00)
Anthropic’s new benchmark finds frontier LLMs can break 65%–86% of tier-1 ciphers — AnthropicAI · 2026-07-29 01:16 (UTC+08:00)
Anthropic posts the papers and chain-of-thought behind its crypto attacks — AnthropicAI · 2026-07-29 01:16 (UTC+08:00)

[source] Anthropic says Claude found new weaknesses in HAWK and round-reduced AES — AnthropicAI · 2026-07-29
[source] Anthropic’s new benchmark finds frontier LLMs can break 65%–86% of tier-1 ciphers — AnthropicAI · 2026-07-29
[source] Anthropic posts the papers and chain-of-thought behind its crypto attacks — AnthropicAI · 2026-07-29
Anthropic’s Claude Mythos reportedly found post-quantum crypto flaws humans missed — Polymarket · 2026-07-29
Anthropic Shows Claude Can Discover Mathematical Flaws in Cryptographic Algorithms — moyix · 2026-07-29
Anthropic’s Claude Mythos AI reportedly finds new attacks on cryptographic algorithms — luisdans · 2026-07-29
Anthropic's Mythos Preview Model Accelerates AES Attack Speed by 200-800x in One Week — tszzl · 2026-07-29