AI cryptanalysis

Where things stand

The topic opens with a report by the New York Times that Anthropic's Claude Mythos Preview model, during testing, found previously unknown attacks against deliberately weakened cryptographic algorithms. According to that report, algorithms of this kind are the sort that secure online financial transactions and private communications, which is what gives the finding its weight: it is the first concrete claim in this thread that a general-purpose AI model produced original cryptanalytic results rather than reproducing known ones. What remains open is how far the result carries. The attacks described apply to weakened algorithms rather than to the full-strength encryption in everyday use, and beyond the New York Times account there is so far no independent technical verification, no published detail of the attacks themselves and no response from cryptographers or standards bodies. Anthropic has not, on the basis of the available reporting, said what the discovery implies for deployed systems. The next steps to watch are whether the underlying work is published in reviewable form and whether outside researchers can reproduce it.

Cryptanalysis is the practice of finding weaknesses in encryption schemes, work that has traditionally required specialist human researchers and long timelines. The algorithms at issue underpin routine digital life, protecting bank transfers, messaging and other private communications, so any new attack against them is treated as consequential. Researchers commonly test methods first against deliberately weakened or reduced versions of algorithms, which are easier to break and serve as a gauge of how strong a new technique is. This topic tracks claims that AI models can perform such analysis on their own, together with the question of whether those claims hold up under independent scrutiny.