Anthropic claims Claude found flaws in cryptography standards

Anthropic announced that Claude identified previously unknown weaknesses in cryptographic algorithms submitted to international standardization bodies. The company's Frontier Red Team published the findings in a paper titled "Discovering cryptographic weaknesses with Claude," positioning the work as evidence that AI can accelerate security research by catching flaws before attackers do.

Human cryptography experts reviewed and validated Claude's findings, according to Anthropic. The company emphasized that the affected schemes were research proposals under evaluation, not encryption systems already deployed at scale.

What Claude actually found

The targets were cryptographic constructions submitted for international standardization, not mainstream encryption like AES or RSA. Anthropic was careful to distinguish between research-stage proposals and battle-tested systems. "The goal is to strengthen cryptography by finding weaknesses before attackers do," the company said.

Standardization processes exist precisely because new cryptographic designs need stress-testing. Claude's contribution, if the claims hold, is speed: an AI model that can sift through mathematical structures faster than human analysts could accelerate the evaluation pipeline.

Anthropic did not disclose which specific schemes were affected or the nature of the vulnerabilities. The company also did not publish timelines showing how long Claude took versus a human team, which makes the efficiency claim harder to evaluate.

A benchmark for AI cryptanalysis

Alongside the paper, Anthropic released CryptanalysisBench, a benchmark designed to measure how well AI models can break or analyze cryptographic primitives. The company hopes the benchmark will help researchers track whether frontier models are becoming more capable at finding security flaws.

Benchmarks matter because they let third parties verify claims independently. If other labs run their models against CryptanalysisBench, the field gets a clearer picture of which systems are actually useful for this work and which are not.

The dual-use problem

Anthropic acknowledged the obvious tension: a model good at finding cryptographic weaknesses could help defenders, or it could help attackers. The company framed this as a reason for continued evaluation of frontier models' cyber capabilities, not a reason to stop the research.

The framing is optimistic. Responsible disclosure works when the party finding the flaw wants to fix it. If these capabilities become widespread, the assumption that defenders will always find bugs first becomes harder to maintain.

“These results suggest that frontier AI systems are becoming increasingly capable of finding previously unknown cryptographic weaknesses.”

— Anthropic

What this does not change

Mainstream cryptography remains intact. Anthropic explicitly said the affected schemes were not widely deployed. The company's statement that "modern cryptographic systems remain highly secure" applies to the protocols you actually use: TLS, Signal, your password manager.

The announcement is less a breakthrough than a proof of concept. Cryptanalysis has always required deep mathematical intuition. If Claude can contribute meaningfully at that level, it suggests a path where AI models become genuinely useful research partners in security work, not just code generators.

Whether that path leads to safer systems or faster-moving adversaries depends on who deploys these tools first and how.

Source: Tech-Economic Times / ET