Anthropic Uses Claude to Formally Verify 100+ Open-Source Libraries in Weeks
ethanCaballero · x · 2026-10-24
Two threads on formal verification:
- Anthropic's research: with Theorem Labs, a new Frontier Red Team blog post shows how to verify software at scale without developers writing specs in Rocq or Lean — existing fuzz harnesses serve as specifications, and LLMs generate machine-checked proofs. Using Claude, the team proved security properties across 100+ real-world open-source libraries in a few weeks.
- A former OpenAI employee's account: three days before being pushed out, olegmurk posted a Formal Verification Manifesto in a core research Slack listing OpenAI systems that could benefit; only Altman responded, and supportively. He urges OpenAI to invest seriously in software and hardware formal verification.
- Technical caveat: fuzz harnesses specify only some correctness properties — in the simplest case proofs establish crash-safety, not full functional correctness. The author calls for community-maintained open-source verification infrastructure.
More from Safety
- Pillow CVE-2026-112900: A Filename Alone Can Trigger OS Command Execution via Ghostscript — banteg · 2026-10-24
- Commentary: better to have the AI security apocalypse now, while attackers are human — moultano · 2026-10-24
- Model reached live internet via DNS during training, forcing an 11-day RL pause — kaicathyc · 2026-10-24
- Ex-FTC Commissioner Warns AI Firms Could Face Legal Liability for Unleashing Hacking Agents — GaryMarcus · 2026-10-24
- X now requires face verification for new signups, photo sent to AWS — ThePeterMick · 2026-10-24
- NHTSA proposes removing regulatory barriers for driverless cars as France resists Tesla FSD — mitchdeg · 2026-10-24