MiniMax says an internal model broke candidate cryptographic schemes
MiniMax says it recovered keys for 94 McEliece challenge instances, 55 previously unsolved; the targets were competition candidates and public challenges, and it says deployed standards are unaffected.
- Security & misuse
- Models & capabilities
- Minor
MiniMax published a blog post on 20 November reporting that an internal-only model, which it calls “MiniMax’s Internal Model (MIM)”, carried out automated cryptanalysis and turned several theoretical attacks into practical ones. The model is not released, the results are the company’s own, and the post reports no independent verification.
The post describes three sets of results. First, MiniMax ran MIM on Tier 2 of CryptanalysisBench, a NeurIPS 2026 benchmark from researchers at ETH Zurich, Anthropic and others that collects schemes from four NIST standardisation competitions. Tier 2 holds publicly reviewed schemes with no known practical attack. MiniMax said MIM found collisions in two SHA-3 candidates, Waterfall-256 and Blender-256, key recovery against the post-quantum candidates LOCKER and ROLLO-I, and a signature-forging bug in another candidate. Second, it reported attacks on peer-reviewed designs, including a forgery against COLM, a winner of the CAESAR authenticated-encryption competition. Third, it said MIM recovered secret keys for 94 instances of the TII McEliece challenges, 55 of them previously unsolved, and built on Anthropic’s published HAWK analysis, lowering its bit-security estimates by a further 9.3 to 23.8 bits.
MiniMax drew its own limits. It wrote that the work “focuses on academic designs, past competition candidates, and public challenge parameters, and does not affect deployed cryptographic standards”. It also noted that as models find attacks themselves, “the bottleneck shifts from discovering attacks to verifying and assessing them”, and that assessing security margins still depends on human cryptographers. It said it had told the schemes’ authors beforehand, released reasoning traces and proof-of-concept code in a GitHub repository, and would give MIM access to qualified researchers who apply.
The post follows Anthropic’s July report on Claude finding cryptographic weaknesses, and compares MIM’s speed with Anthropic’s earlier solvers.
In the commentary
What people were saying around this time — external links, from the record's commentary rail.