ToolNavs Find Useful AI Tools
Submit Sign in
Back to AI information
MiniMax's MIM Model Lands Practical Attacks on Multiple Ciphers: NIST Candidates Broken for the First Time

MiniMax's MIM Model Lands Practical Attacks on Multiple Ciphers: NIST Candidates Broken for the First Time

AI information • Admin • • 24 views

On November 20, 2026, MiniMax's research team published the cryptanalysis results of its internal research model MIM (MiniMax's Internal Model): the model, used only for internal research, produced the first practical attacks against several modern cryptographic schemes that previously had no known practical attack, covering NIST standardization competition candidates, peer-reviewed constructions, and public cryptographic challenge benchmarks. Its reasoning traces and proof-of-concept code have been open-sourced. MiniMax puts it bluntly: cryptanalysis is a clean yardstick for a model's long-horizon reasoning — an attack script either recovers the key or fails, with no room for ambiguity.

What exactly did it break?

Among the NIST candidates, MIM's results concentrate on SHA-3 and post-quantum candidates. For SHA-3 candidate Waterfall-256, only theoretical analysis existed before and no one had ever constructed colliding message pairs; in a blind run, MIM produced the first practical collision attack, finding collisions across all output lengths in 0.045 seconds on a single core. Blender-256's full-round collision resistance had stood unbroken since submission; MIM found full-round collisions in 0.36 to 2.3 seconds. The two post-quantum candidates LOCKER and ROLLO-I previously had no practical key-recovery attack against full parameters either: MIM shrank LOCKER's solver matrix from 42 petabytes to 58 megabytes and recovered the shared secret in 2 seconds, and recovered an equivalent key for ROLLO-I-128 in about 6 seconds. MIM also turned up more than a dozen unreported implementation flaws in official reference code, including a logical AND mistyped as a bitwise AND and randomness-sampling leakage that had survived several rounds of public review.

Why this deserves to be taken seriously

What deserves attention is not that one algorithm was broken, but that the way of verifying has changed. Cryptanalysis is judged in binary: one wrong step in the derivation or one bug in the code, and the attack fails. That MIM could turn theoretical analysis into working attacks across several unrelated schemes shows the model can now connect mathematical reasoning and engineering implementation into one complete chain. For the cryptography community, security evaluations will have to add a new line to the ledger: whether a scheme can withstand automated cryptanalysis, not just human review. The boundaries matter too: MIM targeted academic designs, past competition candidates, and public challenge parameters, and MiniMax states clearly that this work does not affect deployed cryptographic standards; the findings were shared with the authors of each scheme in advance, following responsible disclosure. In the same week, the European Commission confirmed an AI-assisted cyberattack: AI is rewriting both offense and defense at once, and if evaluation cannot keep up, the defenses will be the first to show gaps.

What to watch next

MiniMax has compiled the reasoning traces from its experiments into technical reports, open-sourced them with the proof-of-concept code in the MiniMax-AI/Research repository, and opened MIM testing applications to qualified researchers. Two things are worth watching next: independent cryptographers' verification of these attacks, and whether MIM moves from an internal research model to a publicly available one.

Recommended Tools

More