
On 20 November 2026, MiniMax published a research post describing work by an internal-only model it calls MiniMax's Internal Model, or MIM. The claim is not that today's deployed cryptographic standards have fallen. It is that the model produced practical breaks, meaning running code that recovered a secret or forged a result within a tight compute budget, on several historical NIST competition candidates and on at least two schemes that had passed peer review. It also reports much faster records on public hardness challenges.
Those results, if independently confirmed, would matter for how competitions and journals review designs. They would not, on MiniMax's own wording, change the algorithms now used to protect ordinary internet traffic. The model itself is not released. Access is offered to qualified researchers by email. The post points to technical reports and proof-of-concept code at a MiniMax-AI GitHub repository. When this desk opened that repository, the public page did not visibly list those artifacts.
[1]MiniMax says it evaluated MIM on Tier 2 of CryptanalysisBench, a NeurIPS 2026 benchmark it attributes to researchers from ETH Zurich, Anthropic, and collaborators. Tier 2, as described here, holds 135 publicly reviewed schemes with no known practical attacks, plus 7 production-strength challenge schemes. This desk has not independently opened that paper.
The post gives named examples with timings on a single core. On Waterfall-256, a NIST SHA-3 candidate, it says prior work was theoretical and that MIM found collisions across all output lengths in 0.045 seconds. On Blender-256, another SHA-3 candidate, it says full-round collisions appeared in 0.36 to 2.3 seconds, and that earlier models had mainly used bugs in the C implementation. On LOCKER and ROLLO-I, post-quantum candidates, it reports recovering secrets in about 2 seconds and 6 seconds after work on the underlying problems, including, on LOCKER, shrinking a solver matrix from 42 petabytes to 58 megabytes.
A second cluster is implementation, not design. MiniMax says MIM found more than a dozen unreported bugs in Tier 2 reference code. Examples include a verification typo in DRS that allowed forgeries from the public key alone, randomness leakage in RQC-128, and a decryption-failure path in LAC-128 that hashed uninitialized memory. Those are serious if true, and they are a different kind of finding from a break of the mathematics.
The post then leaves old competition leftovers. It says MIM built a forgery on COLM, a CAESAR authenticated-encryption winner later published in the Journal of Cryptology in 2024, using two encryption queries and no decryption queries. It says a key-recovery attack on SBC-VOLE, a post-quantum signature scheme from ASIACRYPT 2025, recovered the secret key from one known signature in 0.06 seconds across the official parameter sets. On ARADI, a 2024 NSA low-latency cipher, it says practical key recovery now reaches 8 rounds, against a prior public mark of 5, in 2.4 to 16.0 seconds on a single core.
On public challenges, MiniMax says MIM recovered secret keys for 94 TII McEliece instances, including 55 previously unsolved hard instances, and that on instances with prior public records it finished in 1.80 to 5.36 seconds on a single 64-core server, a 6,180x to 43,400x speedup over a CRYPTO 2026 record and over Anthropic's earlier multi-GPU solver. On HAWK, a NIST additional-signature candidate, it says the work lowered bit-security estimates by another 9.3 to 23.8 bits after Anthropic's prior analysis, and recovered an equivalent key in 30 minutes on a single core.
These numbers are unusually specific. They are still vendor numbers until the promised code can be checked.
[1]MiniMax frames cryptanalysis as a clean test for long-horizon reasoning, because a script either recovers the secret or it does not. That is a useful standard. It also raises the bar for the company. The object that can be checked is not the internal model. It is the proof-of-concept code, if it actually appears, and any confirmation from the authors of the named schemes.
[1]MiniMax is clear about scope. The work, it says, focuses on academic designs, past competition candidates, and public challenge parameters, and does not affect deployed cryptographic standards. It says findings were shared with authors in advance.
That sentence should stay attached to every retelling. So should a second limit. Discovery is claimed for an internal model. A repository that does not yet show the promised reports and code cannot carry the weight of the timings.
There is also a category error waiting in the coverage. Implementation bugs in reference code, including typo-level faults, are not the same as a new attack on a design that survived expert review. Both can be real. Mixing them into one line about artificial intelligence breaking cryptography would overstate the design results and understate how often official code is the weak point.
[1]If the named breaks hold up, MiniMax has a point that goes beyond a vendor demo. Practical collisions and key recovery on schemes that had no such attacks would mean automated tools can now finish work that used to stall at a theoretical write-up. COLM and SBC-VOLE, if confirmed, are stronger evidence than a forgotten SHA-3 candidate, because they went through long public review. Faster solvers on McEliece and HAWK challenges would also matter for how parameter choices are justified, even when the scheme itself is not broken in the headline sense.
The bottleneck MiniMax names is verification by human cryptographers, not discovery. That claim is testable, and it does not require anyone to accept that deployed standards are at risk this week.
[1]The next six months should be about artifacts, not adjectives. Watch whether the cited GitHub repository fills in with reports and code that reproduce the named timings on Waterfall-256, COLM, SBC-VOLE, and the McEliece instances. Watch whether the authors of those schemes, or other cryptographers, confirm, narrow, or reject the claims. And watch whether MiniMax ever puts MIM itself in qualified hands, because a proof-of-concept script and a general cryptanalysis model are not the same object.
[1]