MiniMax Cyber Cryptanalysis Breaks NIST Candidates

MiniMax Cyber cryptanalysis results against NIST cryptographic candidates

MiniMax says its Cyber model turned several purely theoretical cryptographic attacks into working exploits. The company’s MiniMax Cyber cryptanalysis results cover old NIST competition entries, recently published ciphers and public math challenges, and many of the attacks run in seconds on one CPU core.

The work targets academic designs and past candidates. MiniMax states that it does not affect deployed cryptographic standards, and that it told the authors of the affected schemes before publishing.

Why cryptanalysis suits a language model

Cryptanalysis attacks the design of a scheme, not a coding mistake in one program. A model has to reason about the math, then write code that carries out the attack within tight data and compute limits. The script either recovers the secret or it doesn’t, so a wrong derivation can’t hide behind confident prose.

That clear pass-or-fail outcome is why MiniMax treats the field as a test of long-horizon reasoning.

Practical breaks of NIST candidates

MiniMax ran Cyber on CryptanalysisBench, a NeurIPS 2026 benchmark built by researchers from ETH Zurich, Anthropic and others. It pulls entries from four NIST contests (AES, SHA-3, lightweight crypto and post-quantum crypto). Its harder tier covers 135 schemes that outsiders had reviewed without finding a practical attack, plus 7 challenge schemes of production strength.

Several of those fell:

  • Waterfall-256, a SHA-3 candidate, had only theoretical attacks. Cyber produced the first working collision attack, with colliding pairs for every output length in under a tenth of a second on a single core.
  • Blender-256, another SHA-3 candidate, had resisted full-round collisions since submission. Cyber produced full-round collisions in somewhere between a third of a second and just over two seconds.
  • LOCKER and ROLLO-I, post-quantum candidates, had no practical key recovery at full parameters. On LOCKER, the model cut the memory the solver needs from 42 petabytes down to 58 megabytes, and the shared secret came out in 2 seconds. For ROLLO-I-128, a usable secret key took about 6 seconds.

Cyber also found more than a dozen unreported bugs in reference code. In the lattice signature scheme DRS, a mistyped logical AND (&&) written as a bitwise AND (&) let Cyber forge signatures for every parameter set from the public key alone, in under a second. In LAC-128, a failed-decryption branch hashed uninitialized memory into the derived key, which broke chosen-ciphertext security.

Flaws in peer-reviewed designs

Most Tier 2 entries were unstandardized, so MiniMax built an automated framework to test schemes from major cryptography venues and competition winners. Cyber found gaps in their security proofs and turned them into attacks.

  • COLM, a CAESAR authenticated encryption winner published in the Journal of Cryptology in 2024, fell to a forgery that needs no decryption queries and two encryption queries.
  • SBC-VOLE, an ASIACRYPT 2025 post-quantum signature scheme, gave up its entire secret key after Cyber saw one known signature, in 0.06 seconds, for all 6 parameter sets.
  • ARADI, an NSA low-latency memory encryption cipher from 2024, had practical key recovery up to 5 rounds. Cyber exploited algebraic patterns in the cipher’s public constants to reach 8 rounds, with the key falling in 2.4 to 16 seconds. It also set new theoretical records on rounds 9 through 12.

Faster solvers for hard problems

Where a scheme has no design flaw, security rests on hard problems from code-based, lattice-based and multivariate math. Public challenge benchmarks help set parameters for NIST post-quantum standards, so MiniMax pointed Cyber at those.

On the TII McEliece Challenges, which relate to Classic McEliece, Cyber developed an algebraic method that makes the equation systems up to 6.59 times smaller. It recovered secret keys for 94 instances, 55 of them previously unsolved. Where earlier records existed, a single 64-core server needed between 1.8 and 5.4 seconds. That works out to speedups of roughly 6,180x to 43,400x over the CRYPTO 2026 record of 9.20 hours and over Anthropic’s earlier multi-GPU cluster solver.

HAWK, an ASIACRYPT 2022 design and a Round 3 candidate in NIST’s additional signatures process, was already hurt. Anthropic had earlier cracked the underlying lattice isomorphism problem for HAWK-256 on a 96-core server, taking 3 hours and 42 minutes, and the authors then pulled out of the competition. Cyber refined both the algorithm and the code. It cut the bit-security estimates for every HAWK parameter set by a further 9.3 to 23.8 bits, and an equivalent secret key took 30 minutes on one core. That is about a 690x gain in core-hours.

What happens next

MiniMax argues the bottleneck is moving from finding attacks to checking them. Working exploit code shows an attack succeeds, but judging the security margin of a design still takes human cryptographers.

The company has released validated reasoning traces in a public repository along with proof-of-concept code. It is also taking applications from cryptography and security researchers for testing access to the model, so outside teams may soon check these claims themselves.

Source: MiniMax

Written by

Saksham Bhargava

Saksham Bhargava runs Alteroid and has been writing about consumer technology for over five years, covering iPhone, Android, Windows, AI tools and gaming. With an MCA and six years of hands-on web development experience, he tests the features he writes about on his own devices and focuses on clear, practical guides that solve real problems. Away from the screen, he enjoys travelling to places most tourists skip.