MiniMax Says an Unreleased AI Model Found Attacks That Cryptographers Had Missed

MiniMax says one of its unreleased AI models broke several encryption designs that experts had previously only been able to attack in theory. The company published the report Nov. 20, 2026, and calls the model MiniMax's Internal Model, or MIM. Nothing in the work has been checked by anyone outside the company.
MiniMax's stated reason for working on cryptography is that the field leaves no room for a plausible-sounding answer: the model has to write code that carries out the attack within set compute budgets, and the script either recovers the secret or it does not. If the account holds, a language model found workable attacks on designs that human cryptanalysts had examined and left standing. The company says the work covers academic designs, past competition candidates and public challenge parameters, and does not affect the cryptographic standards in use today.
The result MiniMax leads with is on Waterfall-256, an entrant in the NIST competition that produced the SHA-3 hash standard. MiniMax's post says earlier work there was purely theoretical, with no pair of messages sharing a hash ever constructed, and that MIM found such pairs at every output length in 0.045 seconds on one processor core. On LOCKER, a candidate from NIST's post-quantum competition, the company reports shrinking the matrix its solver needs from 42 petabytes to 58 megabytes and recovering the shared secret for every parameter set in 2 seconds.
The schemes come from the second tier of CryptanalysisBench, a published benchmark of cryptanalysis tasks drawn mainly from four NIST standardization competitions. MiniMax says that tier holds 135 publicly reviewed schemes with no known practical attack, and that it followed the benchmark paper's own setup. The benchmark is independent of MiniMax — the run reported on it is not. No outside group has repeated any of the results, no independent evaluation of the model has been published, and the model itself cannot be examined.
According to the post, MiniMax sent its findings to the schemes' authors before publishing and has released the model's reasoning traces as technical reports with proof-of-concept code.
Sources
- www.minimax.io
- arXivPreprint
