AI Grinding Paper: Agents Automatically Break 8 Cryptographic Schemes
matthew_d_green · x · 2026-08-25
Lukasz Olejnik et al. propose 'AI Grinding,' a workflow using AI agents to autonomously generate, test, and verify cryptanalytic attack hypotheses.
Core Mechanism:
- Agents cycle through testing hypotheses, filtering viable attack paths, and presenting reproducible candidates with exact witnesses and code to human researchers.
- Humans only review the final evidence.
Findings:
- The paper identifies two common failure modes: algebraic mappings exposing relations and simulator distribution mismatches.
- The workflow successfully broke 8 published cryptographic constructions, including a Ring-LWR commitment scheme and an updatable encryption scheme.
More from Safety
- Insurers Retreat as AI Agents Become 'Uninsured Liabilities' — hbouammar · 2026-08-25
- UK Sovereign AI Fund hires partner to unlock £100M gov procurement — HZoete · 2026-08-25
- Data center debate highlights wasteful practices and policies — _akpiper · 2026-08-25
- Fake Microsoft SysScan Scam: Rigged Scores Force AV Uninstallation — TechNadu · 2026-08-25
- Nvidia employee charged with smuggling advanced chips into China — iamKierraD · 2026-08-25
- AI labs are bad at software engineering; the main threat is incompetence, not powerful models — gerardsans · 2026-08-25