Claude Found New Weaknesses in Two Cryptographic Schemes

Anthropic has presented the results of a study in which a preview version of the Claude Mythos model was used to search for weaknesses in cryptographic algorithms. Cryptography underpins the protection of digital data: some algorithms make it possible to verify identity and the authenticity of messages, while others encrypt information so that unauthorized parties cannot read it. During the experiment, the model discovered previously unknown ways to attack the HAWK digital signature scheme and a specially weakened version of the AES cipher.

The first result concerns HAWK, an experimental digital signature scheme designed, among other things, to withstand the threat posed by future quantum computers. A digital signature mathematically confirms that a message or document was indeed sent by the owner of a particular key and was not altered without detection. According to Anthropic, HAWK had been analyzed by experts for two years, but Claude Mythos Preview took about 60 hours to find a previously unknown attack that halves the scheme’s claimed level of cryptographic security.

The second study focused on a reduced version of AES, a widely known symmetric encryption standard. In symmetric encryption, a single secret key is used both to transform data into an encrypted form and to subsequently restore it. Full AES has been studied by cryptographers for decades and ranks among the most thoroughly tested algorithms. However, the experiment used not the full standard, but an intentionally weakened research variant. In one week, Mythos Preview found a way to speed up an attack on this variant by a factor of 200–800.

Neither result means that the security systems in use today have been compromised. HAWK remains a proposed scheme and, according to Anthropic, has not been deployed anywhere. The work on AES concerns a reduced version of the algorithm and does not make it possible to break its full configuration. The research therefore poses no immediate practical threat to existing services, devices, or encrypted data.

A distinctive feature of the experiment was the model’s high degree of autonomy. Anthropic says that Mythos Preview performed most of the research work autonomously and received only occasional hints from a human. At the same time, the computational cost was substantial: each of the two results incurred approximately $100,000 in API expenses. This level of expenditure currently limits widespread adoption of the approach, but it also demonstrates how much complex intellectual work can be accomplished when a model is given sufficient resources.

Before publication, Anthropic notified the authors of the algorithms under study, as well as its partners in the U.S. government. This disclosure process allows developers and relevant organizations to examine the identified weaknesses before the technical details become publicly available. The company published separate materials fully describing the attacks on HAWK and reduced AES, along with a related record of the model’s chain of thought for the AES study.

The authors view the results as evidence that today’s leading models are capable of conducting certain types of cryptographic research at an expert level. The practical defensive value of such systems may lie in providing additional scrutiny of new algorithms before they are deployed. A model can explore research hypotheses, identify unconventional connections, and help experts detect flaws in the mechanisms on which the confidentiality of online communications and data depends.

At the same time, this capability is dual-use in nature: a tool capable of finding vulnerabilities so they can be fixed could potentially also be used offensively. Anthropic’s announcement itself focuses on research and defensive applications, while the lack of impact on existing systems reduces the immediate risk posed by these two studies in particular. Nevertheless, the results show that evaluating AI capabilities in cryptanalysis is becoming a security task in its own right.

To measure these capabilities more systematically, Anthropic, together with researchers from ETH Zurich, Tel Aviv University, and the University of Haifa, created CryptanalysisBench. It is a benchmark suite designed to test how successfully large language models handle cryptanalysis tasks—the search for weaknesses in ciphers and other protective mathematical constructions. The study moves the discussion of AI in cryptography from isolated demonstrations to measurable tests that make it possible to compare models and track the growth of their capabilities.