Anthropic uses Claude Mythos to find faster attacks on AES and a post-quantum signature scheme

Anthropic published research on July 28, 2026 showing that its gated frontier model, Claude Mythos Preview, discovered improved cryptanalytic attacks against two cryptographic algorithms — the widely-used AES symmetric cipher and HAWK, a post-quantum signature scheme — rather than finding bugs in how they're implemented in code.

What's new

As Anthropic put it, "Using Claude Mythos Preview, researchers at Anthropic have discovered improved ways to attack cryptographic algorithms (the mathematical methods used to keep online data private)." The distinction matters: this isn't a vulnerability in a specific library's implementation of a cipher, it's an improvement to the mathematical attack itself.

For AES, Mythos extended a long-running line of academic research aimed at finding the best possible attack on a reduced, 7-round version of the cipher using a meet-in-the-middle technique — the kind of incremental cryptanalytic progress that typically takes specialist teams months or years to produce. For HAWK, one of the newer post-quantum signature schemes designed to resist attacks from future quantum computers, Mythos produced an attack Anthropic describes as running between 200 and 800 times faster than prior approaches, depending on how the runtime is measured.

Context

Claude Mythos Preview is Anthropic's most capable model to date, released as a gated research preview rather than a general product, with access prioritized for defensive cybersecurity work. It's the same model behind Project Glasswing, Anthropic's initiative to get the model in front of roughly 50 partner organizations to find and fix vulnerabilities in systemically important software — a program Anthropic has said surfaced more than ten thousand high- or critical-severity flaws. This cryptographic-weaknesses research sits alongside that work but targets a different layer of the stack: the algorithms themselves, not the code that implements them.

Why it matters

Most AI-assisted security work to date has focused on finding bugs — memory-safety issues, logic errors, misconfigurations — in existing codebases. Finding improved mathematical attacks on cryptographic primitives is a different and higher-stakes category: a faster attack on a cipher or signature scheme doesn't just affect one piece of software, it potentially weakens every system that relies on that primitive, including standards used across the internet. That an AI model can now meaningfully accelerate that specific kind of research — long considered a slow, expert-driven academic pursuit — is a signal that frontier models are starting to shift the pace of cryptanalysis itself, not just application security. It also raises the obvious dual-use question Anthropic's own gated-access approach is designed to manage: the same capability that helps standards bodies stress-test HAWK before wider adoption could, in less careful hands, be pointed at production systems.