This blogpost from Anthropic describes using the Claude Mythos AI tool to find an attack on HAWK. The attack is not practical at proposed production key strengths, but it improves the best known attacks by quite a lot, and might doom HAWK’s chances in the PQC Signatures On-ramp.
They also found modestly improved attacks against 7-round AES-128 (this reduced-round variant had already been attacked before, and the real AES-128 has 10 rounds), as well as significant problems in KINDI and SPoC, plus a lot of implementation flaws in various algorithms. They attacked hundreds of algorithms in total, some with previously-known flaws and some without.
