MiniMax says internal model broke NIST candidates and sped up cryptanalysis solvers up to 10000x
teortaxesTex · x · 2026-11-22
MiniMax researchers revealed results from an internal-only research model (MIM) used for automated cryptanalysis, claiming previously unknown structural flaws and practical breaks across modern cryptographic schemes.
- Breaking NIST candidates: Evaluated on CryptanalysisBench (NeurIPS 2026, from ETH Zurich/Anthropic et al.) Tier 2 — 135 schemes with no known practical attacks. First practical collisions for Waterfall-256 in 0.045s on a single core; full-round collisions for Blender-256 in 0.36–2.3s; LOCKER solver matrix shrunk from 42 PB to 58 MB with secret recovery in 2s; ROLLO-I-128 key recovery in 6s. Also exploited a dozen unreported implementation bugs (e.g., a && vs & typo in DRS) to forge signatures with zero signing queries.
- Attacking peer-reviewed and competition-winning schemes: working attacks on CAESAR winner COLM, ASIACRYPT 2025's SBC-VOLE, and NSA's 2024 low-latency cipher ARADI (practical key recovery extended from 5 to 8 rounds).
- Faster solvers: on public challenges for classic public-key hard problems, MIM solved dozens of unbroken instances with speedups of hundreds to tens of thousands of times over prior state-of-the-art.
More from Safety
- Ex-NSA's Perlroth: AI is multiplying cyber risk, exploits now beat patches by 7 days — chrisrohlf · 2026-11-22
- Miners 'hacked' Chutes for a year with millions in incentives to forge real end-to-end TEE privacy — const_reborn · 2026-11-22
- Yudkowsky endorses claim he never said p(doom)=99%, but his 2022 words said "probability roughly 1" — teortaxesTex · 2026-11-22
- Second OpenRouter Leak in a Month: MiniMax-Route Prompts May Have Been Read — const_reborn · 2026-11-22
- Security Circle Mocks Execs Blaming Cyberattacks on Chinese Open-Weight Models — nptacek · 2026-11-22
- Many Etsy sellers close up shop over the site's new AI policy, NYT reports — nordicinst · 2026-11-22