MiniMax Says Its AI Broke 4 Old NIST Crypto Candidates

MiniMax, the Chinese AI lab, says an internal-only model broke the designs of four candidate algorithms from past US NIST cryptography competitions, each in seconds. It made the claims in a research post dated 20 November.

MiniMax says the work “does not affect deployed cryptographic standards”. The schemes it attacked are academic designs, past competition candidates and public challenge problems, not the encryption in use today.

The tests come from CryptanalysisBench, a benchmark of competition schemes. We track how models score on tests like it on our AI model benchmarks page.

Business Pill: The Cost of Checking

An evergreen Business Engineer explainer on the cost of checking work. It teaches the general idea only and does not discuss MiniMax’s research.

The key insight: MiniMax says an internal model broke four old NIST candidate designs in 0.045 to about 6 seconds and found more than a dozen reference-code bugs, while saying none of it affects deployed standards.

What MiniMax Says Its Model Broke

NIST, the US standards agency, picks encryption standards through open competitions. Many candidates are submitted, and most are dropped along the way. CryptanalysisBench, which MiniMax says was built by researchers from ETH Zurich, Anthropic and others, collects 135 of those publicly reviewed schemes with no known practical attack.

MiniMax says its model, which it calls MIM, built the first practical collision attack on Waterfall-256, a candidate in the SHA-3 hash competition, in 0.045 seconds on a single core. It says it found full-round collisions in Blender-256, another SHA-3 candidate, in 0.36 to 2.3 seconds.

Two post-quantum candidates fell too, by MiniMax’s account. On LOCKER it says the model shrank the solver matrix from 42 petabytes to 58 megabytes and recovered the shared secret in 2 seconds. On ROLLO-I-128, which reached NIST’s second round, it recovered an equivalent secret key in about 6 seconds.

By our arithmetic, a matrix of 58 megabytes is about 720 million times smaller than one of 42 petabytes, using decimal units.

MiniMax says its internal model broke the designs of four past NIST candidates in 0.045 to about 6 seconds: Wa
MiniMax says its internal model broke the designs of four past NIST candidates in 0.045 to about 6 seconds: Waterfall-256 and Blender-256 (SHA-3) and LOCKER and ROLLO-I-128 (post-quantum). Company claims, not independently reproduced.

Bugs, Not Just Broken Maths

Many findings were in the code rather than the design. MiniMax says the model found more than a dozen unreported bugs in official reference code. In DRS, a post-quantum signature scheme from NIST’s first round, it says a logical AND typed as a bitwise AND let it forge signatures from the public key alone.

Beyond NIST, MiniMax says the model forged ciphertexts against COLM, a winner of the CAESAR competition, and recovered the full secret key of SBC-VOLE, a post-quantum signature scheme from ASIACRYPT 2025, from one known signature in 0.06 seconds.

On the TII McEliece challenges, it says the model recovered keys for 94 instances, 55 of them previously unsolved. It puts its speed-up at 6,180 to 43,400 times over the previous records.

Candidates Have Been Broken Before

Other candidates have fallen before. In 2022 two researchers broke SIKE, then in NIST’s fourth post-quantum round, in about ten minutes on a single core. The same year another researcher broke Rainbow, a finalist, in about 53 hours on a laptop.

Bar chart: Rainbow was broken in about 53 hours on a laptop and SIKE in about ten minutes on a single core in 2022; MiniMax says its model broke ROLLO-I-128 in about 6 seconds
Rainbow (about 53 hours on a laptop) and SIKE (about ten minutes on one core) were broken by researchers in 2022; MiniMax says its model broke ROLLO-I-128 in about 6 seconds.

MiniMax says its model broke the designs, not just the code: it analysed the maths and wrote working attack code. It says it shared its findings with the scheme authors in advance and has published reports and proof-of-concept code on GitHub.

The Structural Read

As we read it, the claim to test is speed of discovery, not broken encryption: the targets are dropped candidates and challenge problems, which is where breaks are supposed to happen.

As we read it, the published proof-of-concept code is the right test: cryptographers can rerun it, which a benchmark score alone would not allow.

MiniMax Research, 20 Nov 2026

“This work focuses on academic designs, past competition candidates, and public challenge parameters, and does not affect deployed cryptographic standards.”

Three Implications

FOR SECURITY TEAMS MiniMax says the deployed standards are not affected.

FOR SCHEME AUTHORS MiniMax says it shared findings with them in advance.

FOR AI LABS MiniMax says verifying model-found attacks still needs human cryptographers.

What Is Not Established

We have not seen public responses from NIST or the scheme authors. MiniMax’s own conclusion says evaluating the security margins of cryptographic designs “still depends on review by human cryptographers”.

MiniMax has not released the model. It is a separate story from the OpenRouter report about requests forwarded to MiniMax this week.

This Week In The AI Supercycle

A World Without an Order, and an AI Central Bank Without a Successor

This week’s issue: the week AI memory stayed largely sold out while the people holding the shares paid: why the AI Supercycle is not a dot-com replay, read through geopolitics, macro and technology, with the gap between building and financing the widest we’ve recorded.

Read the full issue →

Business Engineer Framework

The AI Supercycle

The Business Engineer map of where AI capability turns into business leverage.

Read the Map of AI →

The Bottom Line

MiniMax says its internal model broke four old NIST candidate designs in seconds and found more than a dozen bugs in reference code. It says none of it affects the encryption standards in use. The next test is whether cryptographers confirm the attacks from the published code.

97,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.

A note on sourcing. Every result here is MiniMax’s own claim, from its research post of 20 November, which we read in full; we have not reproduced the attacks. The SIKE and Rainbow precedents come from the researchers’ papers on the IACR ePrint archive. The count of four design breaks is our count of the NIST candidates MiniMax names; the 720 million times figure is our arithmetic. We found no public response from NIST or the scheme authors as of 21 November.

Sources: MiniMax Research, ‘Practical cryptanalysis with MiniMax’s latest Internal Model’, 20 Nov 2026 · MiniMax-AI/Research on GitHub (Cryptography reports and proof-of-concept code) · Castryck and Decru, ‘An efficient key recovery attack on SIDH’, IACR ePrint 2022/975 · Beullens, ‘Breaking Rainbow Takes a Weekend on a Laptop’, IACR ePrint 2022/214 · FourWeekMBA on the OpenRouter report, 21 Nov 2026

Scroll to Top
FourWeekMBA