MiniMax's internal model breaks several post-quantum candidates

MiniMax's internal model breaks several post-quantum candidates

On November 20, MiniMax published a technical blog post describing the cryptanalysis results of MIM (MiniMax's Internal Model), an internal research model. Against a set of NIST candidate algorithms, peer-reviewed schemes and public challenges, MIM produced practical attacks that no one had managed before, and in some cases turned attacks that existed only on paper into programs that run. The technical report, proof-of-concept code and curated model reasoning traces are all posted in the MiniMax-AI/Research repository on GitHub.

MIM is not open to the public. MiniMax says it will give "qualified researchers and developers" access, and applicants need to apply by email.

Which schemes were broken

The blog groups its results into three categories. The first is candidates from past NIST calls for submissions, most of which dropped out long ago but had never been practically attacked:

  • The SHA-3 candidate Waterfall-256 had a real collision constructed in 0.045 seconds on a single core. For another SHA-3 candidate, Blender-256, a full-round collision took 0.36 to 2.3 seconds on a single core.
  • For the post-quantum candidate LOCKER, the shared key was recovered in under 2 seconds across all parameter sets, with the solving matrix shrunk from 42 PB to 58 MB.
  • For ROLLO-I-128, a post-quantum candidate that reached the second round, an equivalent private key was recovered from the public key in about 6 seconds on a single core.
  • The code of the lattice signature scheme DRS contains a mistyped operator, which makes signatures forgeable.

MIM also found flaws such as seed truncation, macro parameter errors and out-of-bounds memory access in the reference implementations of KCL, LEDAkem, ISAP and others. The blog says more than a dozen of these had not been reported before.

The second category is designs published in recent years. A forgery attack was constructed against COLM, a CAESAR competition winner: it needs no decryption queries, and two encryption queries are enough to produce a valid ciphertext. For SBC-VOLE, a post-quantum signature presented at ASIACRYPT 2025, a single known signature is enough to recover the private keys for all six parameter sets, in 0.06 seconds. For ARADI, the low-latency memory encryption cipher the U.S. National Security Agency published in 2024, practical key recovery was pushed from 5 rounds to 8; rounds 9 through 12 remain theoretical.

The third category is hard problems underlying public-key cryptography. In the McEliece challenges run by TII, MIM recovered the keys for 94 instances, of which 55 had never been solved by anyone.

What it is compared against

The blog uses two points of comparison: records presented at the CRYPTO 2026 conference, and the multi-GPU solver Anthropic previously made public. MiniMax says that for McEliece instances with public records, a 64-core server took 1.80 to 5.36 seconds, against the 9.20 hours recorded at CRYPTO 2026, a speedup of 6,180 to 43,400 times. For HAWK, a third-round candidate in NIST's additional signature call, the estimated security level was lowered by 9.3 to 23.8 bits, and an equivalent private key was recovered in 30 minutes on a single core, which MiniMax puts at about 690 times Anthropic's earlier work in core-hours.

All of these multiples come from MiniMax's own report. The blog does not break down how much of the speedup comes from new ideas proposed by the model and how much from engineering and hardware differences, and whether the baselines used the same parameters can only be confirmed from the details in the technical report. Several labs have publicized results from using AI for security research over the past year; this time MiniMax wrote Anthropic's results directly into its comparison table, measuring speed on the same class of problems.

"does not affect deployed cryptographic standards"

Scope and disclosure

MiniMax draws the boundaries fairly clearly: the subjects are academic designs, past competition candidates and public challenge parameters. The list of attacks includes no standardized, already-deployed algorithms such as ML-KEM or ML-DSA. The blog also says that evaluating the security margin of a cryptographic design "still depends on review by human cryptographers." In line with responsible-disclosure practice, the team sent its findings to the authors of each scheme beforehand.

Several questions remain unanswered: MIM's architecture, parameter count and training data have not been disclosed; the results have no independent third-party verification; and the full list of the dozen or so implementation flaws has not been published. With the verification code now open source, the first people able to render a judgment will be the original authors of the named schemes.

Sources: MiniMax technical blog, used to check attack timings for each scheme, the 94 and 55 McEliece challenge instances, and the HAWK security-level reduction; the MiniMax-AI/Research code repository; CocoLoop.