AI Model Comparison

Model profile & tracker

MiniMax's AI Model Broke Crypto Schemes: What Happened

MiniMax's internal AI model claimed practical breaks of 7 cryptographic designs in seconds. What it broke, what it did not, and what is still unverified.

Verified 2026-11-22

On 20 November 2026 MiniMax Research published "Practical cryptanalysis with MiniMax's latest Internal Model", a research post claiming that an internal-only AI model — the post calls it MiniMax's Internal Model, or MIM — found previously unknown structural flaws in academic cryptographic designs and turned several theoretical attacks into practical breaks.

This page separates what the post actually claims from what it does not, the same treatment we gave the Riemann hypothesis AI claim. The short version: real working attack code was released, the claims are MiniMax's own until outside cryptographers review them, and nothing in the work touches the encryption your bank or your messages use.

What MIM broke, in numbers

The post covers three areas: historical NIST standardization competitions, recent peer-reviewed designs, and the classical hard problems behind public-key cryptography.

Target What it is Previous best MIM's result
ARADI NSA 2024 low-latency memory-encryption cipher Practical key recovery on 5 rounds Full key in 2.4–16.0 seconds, 8 rounds, single core; new theoretical records for 9–12 rounds
Waterfall-256 NIST SHA-3 hash candidate Collisions purely theoretical, none ever built First practical collision, all output lengths, 0.045 seconds
Blender-256 NIST SHA-3 hash candidate Only 64-bit implementation bugs exploited Full-round collisions in 0.36–2.3 seconds
COLM CAESAR competition winner (Journal of Cryptology 2024) No practical attacks in the standard model Forgery with no decryption queries; valid ciphertext in two encryption queries
SBC-VOLE Post-quantum signature scheme (ASIACRYPT 2025) No prior practical attacks Full secret key from one known signature in 0.06 seconds, all 6 parameter sets
TII McEliece Challenges Challenge instances based on Classic McEliece CRYPTO 2026 record: 9.20 hours 94 instances solved (55 never solved before), 1.80–5.36 seconds on one 64-core server — a 6,180× to 43,400× speedup
HAWK NIST post-quantum signature candidate Anthropic: HAWK-256 in 3h 42m on a 96-core server Equivalent key in 30 minutes on a single core, ~690× fewer core-hours; bit-security lowered another 9.3–23.8 bits

The McEliece and HAWK rows are the ones that made headlines, because both beat published records — including records set by Anthropic's own crypto-analysis work — by three to four orders of magnitude. HAWK's authors had already withdrawn it from the NIST competition after Anthropic's 2026 analysis; MIM pushes its security margin down further.

What the claims are worth

Three things separate this from a benchmark press release:

  1. Working code was released. Each break comes with proof-of-concept attack code and validation reports, published after the findings were shared with the schemes' authors in advance — coordinated disclosure, not a surprise drop.
  2. The model is internal-only. MIM is not on MiniMax's API, has no price, no card in our model index, and no announced release. You cannot run it, and neither can we.
  3. Nothing deployed is affected. The post says it explicitly: the work targets academic designs, past competition candidates and challenge parameters, not deployed standards. TLS, AES, RSA as used today are outside its scope.

The honest caveat is the same one that hangs over every AI-solves-math claim this year: the results are MiniMax's own measurements of MiniMax's own model. The post itself says the bottleneck has shifted from finding attacks to verifying them, and that assessing security margins "still depends on review by human cryptographers." That review had not happened when this page was written. Until it does, treat every number above as a claim with code attached, not as settled fact.

Does this affect your encryption

No. The post's targets are research designs and withdrawn or unselected candidates. The winners of the standardization processes — AES, SHA-3, and the NIST post-quantum selections that survived review — are not in the break list. ARADI is the closest to deployment of anything attacked, and the practical break covers 8 of its rounds against design criteria that were never published in full.

If you work in cryptography, the technical reports and PoC code are the story: they are reproducible and specific. If you use encryption, the story is the trend line — two AI labs in one year producing publishable cryptanalysis — not any single row in the table.

Why it matters for everyone else

For a site that compares AI models, this result matters the way the Riemann hypothesis claim does: it is evidence about what frontier-scale models can do to problems that were thought to need human specialists, gathered on a problem with an unfakeable answer — the attack code either recovers the key or it does not.

It also widens a gap that our price table cannot show: the most capable analysis models are increasingly not for sale. MIM, like Anthropic's cryptanalysis work, ran on internal infrastructure. The models you can buy by the token are one step behind the models that get pointed at open problems.

FAQ

Is my encryption broken? No. The post targets academic designs and past competition candidates, and states that it does not affect deployed cryptographic standards.

Can I use MIM? No. It is an internal-only research model with no public access, no API listing and no announced release date.

Who verified the claims? As of 22 November 2026, no independent verification had been published. The proof-of-concept code and validation reports are public, so outside cryptographers can check them; the post itself says human review remains the bottleneck.

Does this make MiniMax's commercial models better? No connection is claimed. MIM is a separate internal research model; MiniMax's public lineup in our model index is unaffected by this research.

What was the biggest speedup? On the TII McEliece challenges, MIM solved instances 6,180× to 43,400× faster than the CRYPTO 2026 record, solving in seconds what previously took 9.20 hours.