Vibeleaderboard
← All Intel
Intel / article

Practical Cryptanalysis With Minimaxs Latest Internal Model

Source
minimax.io
Author
MiniMax editorial sitemap
Date

Context

MiniMax says an unreleased research model, which it calls MiniMax's Internal Model or MIM, found working attacks on cryptographic designs that years of human review had left standing (MiniMax's report). The targets are academic proposals, candidates from past NIST standardization competitions and public challenge problems. MiniMax states that none of the work affects deployed cryptographic standards, so the news concerns what a can now do in this field and how well its claims can be checked, with no new risk to the encryption people rely on today.

Cryptanalysis looks for weaknesses in a scheme's design rather than mistakes in someone's code, and MiniMax chose it as a test because the result is all or nothing: the model has to turn its mathematics into a program, and the program either recovers the secret or it does not. The results fall into three kinds that are worth keeping apart. Some are breaks of the design itself: MIM built the first practical collisions for Waterfall-256, a SHA-3 candidate, finding them in 0.045 seconds on one core, forged a valid ciphertext for COLM, a winner of the CAESAR authenticated-encryption competition, with two encryption queries, and recovered the full secret key of SBC-VOLE, a post-quantum signature scheme from ASIACRYPT 2025, from a single signature in 0.06 seconds. Others are bugs in the reference code that submitters published, such as a `&&` mistyped as `&` in the DRS signature scheme that let MIM forge signatures using only the public key. The third kind is faster solvers for the hard problems that public-key schemes rely on. On the TII McEliece challenges, MiniMax reports keys for 94 instances, 55 of them previously unsolved, and solve times of 1.80 to 5.36 seconds on one 64-core server for instances with earlier records, which it puts at 6,180 to 43,400 times faster than both the 9.20-hour CRYPTO 2026 record and a multi-GPU solver from Anthropic.

The work builds directly on Anthropic's. MiniMax evaluated MIM on Tier 2 of CryptanalysisBench, a that researchers from ETH Zurich, Anthropic and other institutions assembled from four NIST competitions, where Tier 2 holds 135 publicly reviewed schemes with no known practical attack. It also extends Anthropic's earlier attack on HAWK, a candidate post-quantum signature scheme. MiniMax says Anthropic's solution took 3 hours and 42 minutes on a 96-core server, while MIM recovered an equivalent secret key in 30 minutes on a single core, about 690 times fewer core-hours, and lowered HAWK's estimated security by a further 9.3 to 23.8 bits across its parameter sets.

Every figure here is MiniMax's own account, the post does not say how many attempts the model made or how many failed, and the model is available only to researchers and developers MiniMax approves. What makes the claims checkable is that MiniMax published technical reports and proof-of-concept code for eight of its targets, among them COLM, SBC-VOLE, HAWK and the McEliece challenges, in its research repository, and says it told the schemes' authors before publishing. A recovered key or a colliding pair of messages can be confirmed by running that code. MiniMax draws the same boundary itself: working exploit code proves an attack succeeds, but judging how much security margin a design has left still falls to human cryptographers.

Key takeaways · AI-distilled
  • MiniMax says its internal-only research model (MIM) produced practical collision attacks on SHA-3 candidates Waterfall-256 and Blender-256, where prior work was theoretical or only exploited implementation bugs, finding collisions in under 3 seconds on one core.
  • Many of MIM's breaks came from reference code, not the math: MiniMax reports over a dozen unreported bugs, including a && typed as & in the DRS signature verifier that let it forge signatures from the public key alone.
  • MiniMax reports MIM forged ciphertexts for COLM, the CAESAR authenticated-encryption winner, with two encryption queries, and recovered the full SBC-VOLE secret key from a single signature, despite both schemes having security proofs or years of review.
  • On the TII McEliece challenges, MiniMax says MIM recovered keys for 94 instances, 55 previously unsolved, and on instances with prior records ran 6,180x to 43,400x faster than the best previous solvers by combining algebraic shortcuts with low-level code optimization.
  • MiniMax limits the claims: the work covers academic designs, past competition candidates and public challenge parameters, not deployed standards, and it argues the bottleneck now shifts from finding attacks to human cryptographers verifying them.
Terms in this piece · Glossary
  • LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.
  • benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
Read the source www.minimax.io
More from MiniMax editorial sitemap
Recommended reads
Comments

Checking sign-in…

Loading comments…