MiniMax's internal model claims first practical breaks on NIST candidates, up to 43,400x faster than Anthropic solvers
Dan_Jeffries1 · x · 2026-11-22
MiniMax published a blog post on its internal-only research model (MIM) for automated cryptanalysis, with PoC code on GitHub:
- First practical breaks: On CryptanalysisBench (NeurIPS 2026, co-created by ETH Zurich and Anthropic researchers) Tier 2, MIM achieved first practical attacks on NIST candidates with no prior practical breaks, including SHA-3 candidates Waterfall-256 (collisions in 0.045s on one core) and Blender-256 (full-round collisions in 0.36–2.3s).
- Post-quantum schemes: LOCKER solver matrix shrunk from 42 PB to 58 MB with secret recovered in 2s; ROLLO-I-128 equivalent key recovered from the public key in 6s.
- Implementation bugs: Over a dozen unreported bugs in official reference code of DRS, RQC-128, LAC-128, KCL, LEDAkem, ISAP and others, enabling signature forgery with zero signing queries.
- Speedups: Up to 43,400x faster than Anthropic's multi-GPU solver on LOCKER (single 64-core server); beat Anthropic's 96-core HAWK attack in 30 minutes on one core; pushed key recovery on NSA's ARADI cipher from 5 to 8 rounds; solved 55 previously unsolved McEliece instances.
All figures are from MiniMax's own report and await independent verification.
Related event: MiniMax's Internal Model Cracks Multiple NIST Candidate Algorithms(4 posts)→
More from Models
- Google testing Inbox mode for Antigravity to shift users into approving AI outcomes — testingcatalog · 2026-11-22
- Can you resell ChatGPT Pro's $200 monthly API credits in a commercial SaaS? — infolily · 2026-11-22
- Insider teases: top AI accounts are early-testing an "absolute monster" model — iruletheworldmo · 2026-11-22
- Inkling 2 mini's 1M-context KV cache dissected: only 8 of 48 layer visits keep full cache — teortaxesTex · 2026-11-22
- MiniMax says internal model cracked multiple NIST candidates with no known attacks — MiniMax 稀宇科技 · 2026-11-22
- GPT-6.2 is hard to steer via system prompt — OpenAI says use it directly in Codex — pvncher · 2026-11-22