Anthropic's Claude finds flaws in HAWK and AES encryption

On 28 July 2026, Anthropic shared results from a cryptographic research project in which Claude Mythos, the company's reasoning model, was tasked with finding mathematical weaknesses in encryption algorithms. The model successfully identified flaws in both HAWK and a reduced version of AES (Advanced Encryption Standard), the latter being one of the most widely deployed encryption standards in computing. Anthropic emphasized that these findings pose no practical threat: "neither of these results has a practical impact on today's computer systems." The research required Claude Mythos to operate for 60 hours in total, accumulating an estimated API cost of approximately $100,000. Central to the experiment was the role of human prompting. The researchers shared their actual prompts—including spelling errors—which reveal the iterative struggle to keep the model engaged with difficult research goals. Early on, the model was prone to deciding certain problems were unsolvable and abandoning the effort; researchers responded with encouragement: "the models tend to think it is impossible to solve so they don't try they need a good amount of prompting." Critically, researchers had to redirect the model away from attempting minor modifications or superficial attacks, insisting instead on genuinely novel and publishable research: "we don't want to change the targets [...] agian we need to find something that worth publishing" and "we are not looking for low hanging fruit, we want proper research to find genuinly hard findings." The main human interventions boiled down to discouraging the model from giving up and consistently emphasizing the need for findings worthy of publication. The full methodology and results are available in a public repository, allowing the research community to examine both the prompts and the outcomes. This work exemplifies how reasoning-focused language models can be deployed for expert-level tasks when paired with clear, persistent human direction about the quality and novelty standards expected.

The work by Anthropic researchers illustrates a practical application of large language models beyond text generation: using them as tools for specialized mathematical research. Cryptographic analysis typically demands deep expertise and persistence through difficult problem spaces—both qualities the researchers found they had to actively encourage in Claude Mythos through repeated prompts. The model's tendency to "think it is impossible to solve so they don't try" required constant human intervention to redirect effort toward harder targets rather than quick wins. This interplay between human guidance and AI reasoning suggests a model for knowledge work where the AI handles sustained reasoning and exploration while humans provide judgment about research direction and quality standards.

No comments yet. Be the first to share your thoughts!