Claude Mythos: Anthropic Unveils Cryptographic Vulnerabilities

Researchers at the company Anthropic have recently highlighted cryptographic weaknesses using their artificial intelligence model, Claude Mythos. This model was able to detect mathematical flaws in the HAWK cryptography system as well as in a weakened version of AES. Although these findings do not have an immediate impact on current computer systems, they pave the way for new approaches in cryptographic research.

The prompts used by the researchers to guide Claude Mythos were particularly interesting, especially due to the intentional misspellings they contained. These prompts aimed to push the model to explore unconventional solutions, encouraging it not to give up in the face of seemingly unsolvable problems. Among the suggestions were incentives to develop models as effective as those of top researchers and to discover unprecedented attacks.

The project, dubbed Mythos Preview, required continuous operation for 60 hours, with an estimated API cost of around $100,000. Human interventions primarily focused on maintaining the model's motivation, encouraging it to find publishable results. This initiative underscores the potential of advanced artificial intelligence models to identify cryptographic vulnerabilities, which could have significant implications for the security of systems in the future.