Claude found 2 flaws in algorithms deemed indestructible—this is a first... and it changes everything.

Claude Mythos Preview continues to uncover vulnerabilities in lines of code. Anthropic’s AI has just uncovered two unprecedented mathematical flaws in encryption systems. This is the very first time that an artificial intelligence has detected theoretical weaknesses in the design of cryptographic algorithms.Claude is now “capable of finding mathematical flaws in the algorithms themselves”, Anthropic celebrates.

New Anthropic research: Discovering cryptographic weaknesses with Claude.

Claude Mythos Preview has helped our researchers find weaknesses in cryptographic algorithms—the mathematical methods that are used to keep data private.

Read more: https://t.co/TYKLjb3Q7V

— Anthropic (@AnthropicAI) July 28, 2026

For the record, Anthropic has made Claude Mythos available to a handpicked group of partners, including the National Security Agency (NSA). Under the Glasswing project, the AI model was deployed across a wide range of infrastructures. It was thus able to uncover a 29-year-old security flaw in Squid, one of the world’s most widespread web proxy servers, as well as a 27-year-old vulnerability in the OpenBSD system. More than 10,000 vulnerabilities have been fixed thanks to Mythos.

Read also: 16 million computers affected – Claude reveals a 9-year-old Linux flaw

A flaw in a post-quantum candidate

First, Mythos examined how HAWK, a digital signature scheme, works. This cryptographic tool makes it possible to prove that a message or document did indeed come from a specific person and has not been altered. Introduced in 2022, HAWK is one of the encryption systems presented as a post-quantum candidate.In plain terms, the protocol must be able to withstand cyberattacks carried out by quantum computers, which are expected to arrive within the next few years.

It was selected as part of a competition run by the US agency NIST (National Institute of Standards and Technology). For the past four years, the agency has been running this competition to identify the best resilient encryption standards in a world dominated by quantum computers. After two years of review, HAWK is among the last remaining candidates in the competition.

However, a major vulnerability lies at the heart of HAWK, and no researcher was aware of it before Claude got involved. Supported by an Anthropic researcher, the AI uncovered a mathematical weakness in the encryption protocol. It took Mythos only 60 hours of work to pinpoint the flaw. This vulnerability can be exploited to more easily guess the secret key used to decrypt all content protected by HAWK.

In theory, a powerful computer could use this breach to attempt to crack the standard. In practice, the Anthropic team was able to recover the secret key on a simple server in a matter of hours. To fix this enormous problem, HAWK’s key sizes would have to be doubled, which would deprive the system of most of its practical advantages. Note that Claude Mythos’s discovery concerns the system’s smallest configuration, HAWK-256, rather than the more robust versions.Nevertheless, the vulnerability affects the very mathematical principle underpinning HAWK’s security, reducing its relevance compared with other algorithms.

Read also: Thank you, AI, Linux has fixed more than 400 security vulnerabilities in 24 hours!

A vulnerability in a reference algorithm

The second vulnerability uncovered by Claude Mythos lies in AES (Advanced Encryption Standard), the encryption system most widely used worldwide to protect digital data, whether files, communications, or passwords. Developed in the early 2000s, the protocol established itself as the global benchmark for securing both government classified data and banking transactions.

At first, Claude dug in its heels, deeming it far too difficult to uncover a vulnerability in how AES works. After three days of encouragement and persistence, Mythos finally developed a new method of attack against Advanced Encryption Standard. Named Möbius Bridge, this tactic makes it possible to carry out an attack against the protocol much faster than in the past. The protocol was not compromised, but its security is now considered weakened. In its report, Anthropic clearly states that the vulnerability uncovered by Claude affects only the reduced version of the algorithm.There is therefore no danger to banking systems, nor to any infrastructure relying on AES.

“None of these findings has any practical impact on current computer systems; no software in production will need to be modified”, says Anthropic, even if these weaknesses could theoretically “put the data of billions of users at risk”.

100,000 dollars to discover a weakness in an algorithm

Finally, Anthropic reveals that each of the two discoveries has cost approximately 100,000 dollars in computing costs. The startup says it scrupulously followed a responsible disclosure procedure for the two vulnerabilities, in line with cybersecurity best practices worldwide. Before publicly unveiling its discoveries, Anthropic coordinated with the algorithm designers and with U.S. government officials.

The American group emphasizes that AI is now capable of “detecting flaws in cryptographic designs that had eluded researchers for years”. To improve the security of algorithms, Anthropic believes it is imperative to scrutinize “modern systems”, which have “not received the scrutiny they deserve”.The start-up expects “significant weaknesses” to be uncovered by AI models in the near future.

👉🏻 Follow tech news in real time: add 01net to your sources on Google, and subscribe to our WhatsApp channel.