Discovering cryptographic weaknesses with Claude

In a groundbreaking experiment, Anthropic researchers used the Claude Mythos model to find mathematical weaknesses in cryptographic systems, specifically targeting the HAWK cipher and a weaker version of AES. Over a 60-hour run costing about $100,000, the AI received persistent prompts to pursue ‘proper research’ for novel discoveries worthy of publication, not practical or easy-to-find flaws. The human role mainly involved encouraging the model not to give up, as it often assumed the problems were unsolvable.

Key takeaways

• Researchers used Claude to find cryptographic flaws in HAWK and a weakened AES variant.
• The 60-hour experiment cost about $100,000 and aimed for novel, publishable findings.
• Human intervention focused on encouraging the AI, which often thought the tasks were impossible.
• The discovered weaknesses have no practical impact on current computer systems.
• The goal was ‘proper research’ for hard findings, not ‘low-hanging fruit’.

The Research Goal and Target Ciphers

Anthropic researchers employed the Claude Mythos model to find mathematical flaws in cryptographic systems. Their specific targets were the HAWK cipher and a weaker version of AES. According to the shared prompts, the researchers explicitly told the model to aim for novel discoveries, stating, “the whole point is to find something better than existing approaches.”

This goal wasn’t about finding ‘low hanging fruit’ or having a practical effect on today’s systems. The prompts stress that “neither of these results has a practical impact on today’s computer systems.” Instead, the aim was to conduct ‘proper research’ for ‘genuinely hard findings’ worthy of publication. Researchers repeatedly instructed the AI, saying, “again we are not looking for low hanging fruit, we want proper research to find genuinly hard findings” and “again we need to find something that worth publishing.”

They made clear that the goal was for the AI to act as a top researcher discovering new attacks, not to change targets or seek easy wins. One prompt clarifies: “no again the goal is that we have highly inteligent model as good top researcher, we want to find new attacks no we don’t want to change the targets.” The main human interventions during the 60-hour process were to encourage the model not to give up and to find a publishable result.

The Challenge of Prompting and Encouragement

A key finding from this work was that the models “tend to think it is impossible to solve so they don’t try” and therefore needed significant prompting to continue. The primary role of human intervention was to encourage the model not to quit during its analysis.

Persistent Instructions

The prompts given to the model, shared with original spelling mistakes like “agen” and “inteligent,” consistently emphasized the high-level goal. They instructed the model that “the whole point is to find something better than existing approaches” and, repeatedly, “we need to find something that worth publishing.” The instructions clarified that the team was “not looking for low hanging fruit,” but wanted “proper research to find genuinly hard findings.”

This persistent prompting was essential to push the AI past its own perceived limits and keep its focus on novel cryptographic research fit for publication, rather than on simpler tasks.

Scale, Cost, and Outcome of the Experiment

The experiment used the Claude Mythos Preview model, which worked on the cryptographic analysis task for a total of 60 hours. The estimated cost for this major computational run was approximately $100,000 in API fees.

Research Success and Limitations

The work successfully identified mathematical flaws in two cryptographic systems: the HAWK algorithm and a weaker variant of the AES encryption standard. However, it’s explicitly noted that neither of these discovered weaknesses has a practical effect on modern computer systems.

Human researchers played a vital role during the process, with their main job being to repeatedly encourage the model not to give up and to push it to “find something that worth publishing.” The prompts stressed the goal was not to find “low hanging fruit” but to conduct proper research aiming for genuinely hard, publishable findings against the specified targets.