# China’s MiniMax says its unreleased AI broke a contest-winning encryption scheme and beat Anthropic’s code-breaking records

> Source: <https://madrobot.blog/2026/11/22/minimax-internal-ai-cracks-ciphers-beats-anthropic/>
> Published: 2026-11-22 14:08:00+00:00

**Key takeaway:** MiniMax's MIM forged a COLM ciphertext with two encryption queries, recovered SBC-VOLE keys in 0.06 seconds and cut a LOCKER attack's solver matrix from 42 petabytes to 58 megabytes, suggesting AI cryptanalysis can turn theoretical breaks into practical ones.

MiniMax’s cover artwork for its cryptanalysis research post. Image: MiniMax

MiniMax, the Shanghai AI company, says an AI model it has not released broke an encryption scheme that survived five years of expert review and beat two of Anthropic’s code-breaking records, in a [research post on Nov 20](https://www.minimax.io/blog/practical-cryptanalysis-with-minimaxs-latest-internal-model). The results are MiniMax’s own and unconfirmed, but it has published papers and working attack code anyone can check.

## A forged message in two tries

The model, which MiniMax calls MIM (short for MiniMax’s Internal Model), is not available to the public. Its headline result is against COLM, an authenticated encryption scheme (the kind that keeps a message both secret and tamper-proof) picked for the final portfolio of the CAESAR competition after five years of public review and specified in the Journal of Cryptology in 2024.

COLM claims 64-bit integrity, and no attack faster than about 2<sup>64</sup> operations was known in the standard setting. MIM found a structural flaw and forged a valid ciphertext for COLM’s main mode with just two encryption queries and no access to decryption, according to MiniMax’s [technical report](https://github.com/MiniMax-AI/Research).

MiniMax lists two more peer-reviewed targets:

- **SBC-VOLE** , a post-quantum signature scheme from ASIACRYPT 2025: the full secret key recovered from one known signature in 0.06 seconds, across all six parameter sets.
- **ARADI** , a memory encryption cipher the NSA published in 2024: practical key recovery pushed from 5 rounds to 8, taking 2.4 to 16 seconds on a single core.

## Old NIST candidates fall in seconds

MiniMax also ran MIM on CryptanalysisBench, a benchmark Anthropic built with academics at ETH Zurich, Tel Aviv University and TU Berlin. Its harder tier holds 135 publicly reviewed schemes from past NIST competitions that had no known practical attacks. MiniMax says MIM broke several:

- **Waterfall-256** , a SHA-3 hash candidate: collisions in 0.045 seconds on one core, where earlier attacks were only theoretical.
- **LOCKER** , a post-quantum candidate: the attack’s solver matrix shrunk from 42 petabytes to 58 megabytes, recovering the shared secret for every parameter set in 2 seconds.
- **DRS** , a post-quantum signature scheme: a typo in the official code, a logical AND (&&) written as a bitwise AND (&), let MIM forge signatures using only the public key in under a second.

The company says MIM found more than a dozen unreported bugs in official reference code along the way.

## Beating Anthropic on its own turf

The comparison with Anthropic is direct. In July, Anthropic [said](https://www.anthropic.com/research/discovering-cryptographic-weaknesses) its Claude Mythos Preview model had found a new attack on HAWK, a post-quantum signature scheme in NIST’s standards process, and the HAWK team withdrew it. Anthropic’s researchers recovered a HAWK-256 key in 3 hours and 42 minutes on a 96-core server.

MiniMax says MIM built on that work and recovered an equivalent key in 30 minutes on a single core, about 690 times less computing time, while cutting HAWK’s estimated security by a further 9.3 to 23.8 bits across all its sizes.

On the TII McEliece Challenges, a public set of key-recovery puzzles based on the cryptosystem behind NIST candidate Classic McEliece, MIM recovered 94 keys, 55 of which no one had solved before. On instances with earlier records, it took 1.80 to 5.36 seconds on one 64-core server, which MiniMax says is 6,180 to 43,400 times faster than both a CRYPTO 2026 record of 9.20 hours and Anthropic’s multi-GPU solver. Anthropic hasn’t commented.

## What MiniMax isn’t claiming

MiniMax says the work covers academic designs, old competition entries and challenge puzzles, and “does not affect deployed cryptographic standards”. It says it told each scheme’s authors before publishing, but none of them, nor any independent cryptographer, has publicly confirmed the attacks yet. The proof-of-concept code is public, so they can be tested.

The company also says machines have not replaced people: assessing how secure a design really is “still depends on review by human cryptographers”. Researchers can apply for access to MIM by email.

Anthropic’s own model has done similar work: Claude [explained the origin of the NSA’s mystery curve seeds](https://madrobot.blog/2026/10/16/claude-nist-ecdsa-curve-seeds-nsa-mystery-solved-lrand48/) last month, and Google DeepMind’s AI recently [proved key parts of Signal’s encryption secure](https://madrobot.blog/2026/11/10/deepmind-alphaproof-code-signal-protocol-proof/). Anthropic has also warned that [a Chinese model anyone can download can build working hacks](https://madrobot.blog/2026/09/29/anthropic-glm-5-3-zai-cyber-exploits-safeguards-open-weight/). MiniMax’s last public release, [M3.1 Flash](https://madrobot.blog/2026/10/20/minimax-m3-1-flash-price-benchmarks-deepseek-half-cost/), competed on price.

## Why it matters

Until now, the strongest AI code-breaking results came from Anthropic, which has limited access to its Mythos model. A Chinese lab now says a model it is also keeping in-house can match and beat that work, with code to prove it. If the attacks hold up, AI-driven cryptanalysis is no longer a one-company field.

*Sources: [MiniMax, Practical cryptanalysis with MiniMax’s latest Internal Model](https://www.minimax.io/blog/practical-cryptanalysis-with-minimaxs-latest-internal-model); [MiniMax-AI/Research on GitHub (technical reports and code)](https://github.com/MiniMax-AI/Research); [Anthropic, Discovering cryptographic weaknesses](https://www.anthropic.com/research/discovering-cryptographic-weaknesses).*
