Nobody Will Read the Lean: Cryptographer Urges High-Level Languages for AI-Generated Proofs
jedisct1 · x · 2026-10-14
Nadim Kobeissi (Symbolic Software) publishes an IACR paper arguing that as AI formal proofs become cheap, the bottleneck is humans who can read what was proved.
Key points:
- In Sept 2026 an AI produced a Lean proof of Fermat's Last Theorem in 11 days; Lean can verify steps but not whether the theorem says what you think—definitions, also AI-written, decide that.
- Precedent: a Lean library's Schnorr signature security theorem passed the checker yet held for every adversary, promising nothing.
- He calls for high-level cryptographic languages with short, owner-written models and publicly agreed, versioned definitions of goals like forward secrecy; on a one-page Verifpal model, two of five verdicts flip under alternative standard definitions.
- Agreeing on definitions takes years of public argument better provers can't shorten, so work should start now.
More from Safety
- Researchers make first contact with a wild Chinese AI agent fleet run by Tencent, targeting Alibaba's Amap — dylfreed · 2026-10-14
- Researcher using GLM gets Microsoft MSRC thanks for 23 Windows CVEs, joins Z.ai — Zai_org · 2026-10-14
- Anthropic warns small AI models could soon self-replicate across IoT devices — Ars Technica AI · 2026-10-14
- Agent Baseline: Docker, Barclays et al. Publish Open Reference Architecture for Securing Autonomous Software Factories — AccBalanced · 2026-10-14
- Alleged OpenAI spy vibe-coded a surveillance device on his wife's chatbot account — The Decoder · 2026-10-14
- Gemini 4 Argon Tops SpeechMap at 87% as Google Stays Consistent on Speech — sebkrier · 2026-10-14