Anthropic's Claude Mythos Finds Flaws in Data Encryption Algorithms

Anthropic announced on July 28, 2026 that its newest research model, Claude Mythos Preview, independently discovered new weaknesses in cryptographic algorithms that protect data, including one of the US National Institute of Standards and Technology's (NIST) candidates for a quantum-resistant encryption standard. The company stresses that the model found mathematical flaws in the algorithms themselves, not just implementation bugs, a feat that has until now taken the world's best cryptographers years of work.

Attack on a NIST Candidate

The first discovery concerns HAWK, a digital signature scheme among the third-round candidates in NIST's competition for quantum-computer-resistant cryptography. Mythos Preview improved on the best previously known attack against the algorithm, effectively halving its key strength on a logarithmic scale, a drop from 2^64 to 2^38 operations needed to break the protection.

HAWK had already undergone two years of expert review across two rounds of the NIST competition, yet Mythos Preview found a more effective attack in just 60 hours of work. Anthropic says the entire experiment cost around $100,000 in model API usage fees. HAWK is not yet deployed in any production system, so the discovery poses no direct threat to currently active protections, but it does mean the algorithm in its current form is not fit for deployment without enlarging its key sizes.

A Faster Attack on AES

The second discovery concerns the Advanced Encryption Standard, the most widely used symmetric cipher in the world, which protects network traffic, drives and messaging apps among other things. Mythos Preview developed a new fingerprinting method called Möbius Bridge, which attacks a version of AES reduced to 7 of its 10 standard encryption rounds and is 200 to 800 times faster than the best previously known attacks of this kind.

According to Anthropic, discovering the method alone took the model a week, with full verification by human researchers taking another month. The company notes that reaching this result required just three substantive prompts spread over three days, during which the model generated more than a billion tokens of analysis. As with HAWK, the attack does not threaten the full, ten-round version of AES-128 used in practice today.

Thousands of Flaws Worldwide

The cryptographic discoveries are only part of a broader picture. Anthropic disclosed that Mythos Preview identifies zero-day vulnerabilities in every major operating system and web browser, and in benchmark tests achieved full control-flow hijacking on ten separate, fully patched targets. Among the specific findings are bugs humans overlooked for decades: a 27-year-old flaw in OpenBSD's SACK TCP implementation, a 16-year-old defect in the FFmpeg library's H.264 codec, and a 17-year-old remote code execution vulnerability in FreeBSD NFS.

The scale and cost of these discoveries are striking. Scanning a thousand OpenBSD repositories cost under $20,000, a single run that led to a critical vulnerability cost under $50, and identifying the FFmpeg flaw came in at around $10,000. That is a fraction of the budgets technology companies currently allocate to human-led security audits.

Because more than 99 percent of the vulnerabilities Mythos Preview found remain unpatched, Anthropic decided not to release the model to the general public. Access has been limited to a narrow group of industry partners and open-source developers under the Project Glasswing program, whose goal is to patch critical software before similar tools fall into the wrong hands.

99 percent of the vulnerabilities we found have not yet been patched, so it would be irresponsible to disclose their details - Anthropic

What It Means for Data Security

For companies and institutions that base their data protection on standards such as AES, TLS or SSH, the news cuts both ways. None of the attacks described breaks the full versions of these algorithms as currently used, but the speed at which an AI model caught up to and surpassed years of expert work suggests that the cryptography standardization process, especially for post-quantum algorithms, may need to revise its assumptions about how much time new algorithms have before someone finds a weakness in them.

Anthropic acknowledges that the same capabilities that let Mythos Preview patch flaws faster than ever could also fall into attackers' hands once similar tools become more widely available. Still, the company expects that in the long run defenders will gain the greater advantage, since they can use AI to audit their own code before it reaches production.

Over the long term, we expect defensive capabilities to dominate - Anthropic

For Polish companies and institutions that will be migrating to post-quantum cryptography standards in the coming years in line with recommendations from EU and national cybersecurity regulators, the takeaways from Anthropic's research are practical. Choosing algorithms still going through NIST's standardization process carries the risk that their parameters will need to change before deployment, and the pace at which AI uncovers such weaknesses may now grow faster than the pace at which they can be patched.