Anthropic published one of those posts designed to make every security lead sit up a little straighter: using Claude Mythos Preview, its researchers say they found a significantly stronger attack on the post-quantum signature candidate HAWK and a faster attack on a reduced version of AES. The company is careful, and correctly so, to note that this does not break production systems. HAWK is still a candidate, the AES result is for a reduced-round variant, and nobody needs to rotate keys because a blog post had a good quarter. Anthropic's write-up is here.\n\nThe spicier number is the one buried in the methodology: Anthropic says each result cost roughly $100,000 in API usage to develop. That is a very 2026 sentence. The frontier-lab status game is no longer just who has the smartest model. It is who can afford to point that model at hard research problems until something publishable falls out.\n\nHacker News, naturally, did not spend all its time on the math. Part of the room focused on whether this meant AI was becoming a serious cryptanalysis partner. Another part immediately fixated on the prompting, because Anthropic included gloriously unglamorous examples of researchers nudging Claude with variations on no, not that, find something worth publishing. The thread is here.\n\nThat is the real industry tell. The discourse has already moved past wow, the model did a thing. Now the argument is about harnesses, iteration, oversight, and whether prompt engineering is a real profession or just project management wearing a headset. If you squint, Anthropic's post is both a research result and an accidental status update on AI work itself: the machines are getting sharper, the humans are still steering, and the meter is very much running.