Researchers at Anthropic employed the Claude Mythos model (Mythos Preview) to search for mathematical flaws in the HAWK post‑quantum signature scheme and in a deliberately weakened variant of AES. Anthropic says that neither finding has a practical impact on today’s computer systems.
How the work was done
The team guided the model through iterative, human‑in‑the‑loop prompting. The prompts shared by Anthropic — including spelling mistakes — illustrate that the model often assumes a problem is impossible and may stop exploring unless encouraged. According to the disclosed prompts, researchers repeatedly prompted the model to continue and to "find something that worth publishing."
Mythos Preview ran for about 60 hours in total, and Anthropic estimates the associated API cost at roughly $100,000. The main human interventions reported were motivational: urging the model not to give up and to pursue publishable directions.
What was found and why it matters (and why it doesn’t pose immediate danger)
The reported issues are theoretical: mathematical weaknesses in HAWK and in a weakened AES variant. Anthropic emphasizes these do not translate into an immediate practical threat against current systems, i.e., they are not describing an exploitable, in‑the‑wild attack.
Significance of the experiment
The exercise demonstrates that large language models, when steered by careful prompting and human oversight, can assist in theoretical research such as cryptographic analysis. It also highlights limitations: models may prematurely conclude problems are unsolvable or produce unhelpful output, so active human guidance and validation are necessary.
Public discussion
Details of the project and the shared prompts attracted attention on platforms including Hacker News. The disclosed runtime and cost figures, together with the prompts, have prompted discussion about how large language models should be used responsibly and effectively for security research.
(This article is based on publicly shared information and does not add independent technical verification or new numerical data.)
