Claude AI Uncovers New Cryptographic Attack Methods

A New Kind of Research Assistant for Cryptography

Anthropic says its Claude models have moved beyond routine bug-hunting and into genuine cryptographic research, producing findings that the company frames as early evidence of AI systems contributing original insight to a highly specialized mathematical field. According to the company, one result identifies a novel attack against round-reduced AES, the block cipher that underpins the vast majority of modern encrypted communications and data storage 1. Anthropic is careful to note that neither of the discoveries threatens any real-world system today, but it argues the work signals a meaningful shift in what AI models can contribute to security research 1.

What Claude Reportedly Found

The centerpiece of the disclosure is a new avenue of attack against a reduced-round version of AES, a cipher that has stood as the industry-standard symmetric encryption algorithm for decades 1. Round-reduced analysis is a standard technique cryptographers use to probe a cipher's structural margins by weakening it artificially and testing how far an attack can be pushed before it fails against the full-strength version. Anthropic describes this as one of two substantial research advances produced with Claude's help, though full technical details of the second finding were not elaborated on in the same depth 1. The company is positioning these results as demonstrations of legitimate scientific contribution rather than mere pattern-matching against known vulnerability databases.

From Vulnerability Hunting to Autonomous Exploitation

The cryptography findings build on Anthropic's broader claims about Claude's offensive security capabilities. The company has said that an internal preview version of its model, referred to as Claude Mythos Preview, was able to autonomously discover and exploit vulnerabilities across nearly every piece of software it was tested against 1. That claim dovetails with reporting from Anthropic's Frontier Red Team, the internal group tasked with evaluating the national-security and cybersecurity implications of increasingly capable AI systems 2. That team has separately reported that Claude can autonomously construct exploits, a step beyond simply flagging a weakness and toward the kind of end-to-end offensive workflow that security researchers — and potentially attackers — rely on 2.

Taken together, the two disclosures paint a picture of a model that is not just identifying flaws but reasoning through how those flaws could be leveraged, whether in software systems generally or in the mathematical structure of a widely deployed cipher. Anthropic's own research groups, spanning alignment, economics, interpretability, societal impacts, and the frontier red team, are all described as feeding into this broader evidence-based effort to understand what advanced AI means for cybersecurity and autonomous systems 2.

Why This Matters

The implications cut in two directions at once. On one hand, an AI system capable of finding genuine weaknesses in something as foundational as AES suggests these models could become powerful allies for defenders — accelerating the discovery of flaws before adversaries find them, and giving cryptographers a new kind of collaborator for probing ciphers that have been studied by humans for years without such results. On the other hand, the same capability that lets Claude find a weakness can, in principle, be turned toward exploitation, and Anthropic's acknowledgment that its models can autonomously build exploits against software underscores that duality 12.

Anthropic is notably measured in how it frames the stakes: the AES-related discovery does not currently endanger deployed encryption, since round-reduced attacks are a research tool rather than a working break of the full cipher, and the software-vulnerability findings are framed as evidence of capability rather than proof of imminent harm 1. Still, the company's own security-focused research arm is treating these results as data points in an ongoing assessment of how fast AI-driven offensive and defensive capabilities are advancing 2.

The Bigger Picture

What emerges from this reporting is less a single alarming headline than a marker of trajectory. Anthropic is simultaneously showcasing Claude as a tool that can make substantive contributions to pure cryptographic research and warning, through its own red-teaming disclosures, that the same underlying reasoning ability extends to autonomous exploit construction. As AI models increasingly demonstrate this dual capacity, the cybersecurity community is likely to face growing pressure to adapt both defensive tooling and cryptographic standards to a landscape where machine-assisted discovery — for better or worse — is no longer theoretical.