An AI auditor found a critical bug in OpenVM, but only with the right crypto context

Security firm zkSecurity has published a case study in which an AI auditor found a critical soundness bug in OpenVM, a zero-knowledge virtual machine, and the details are a useful reality check on what these tools can and cannot do. Off-the-shelf prompting failed. Claude Opus 4.6 and 4.7 and Codex 5.3 and 5.4, run with standard prompts and expert-written skills, did not surface the flaw. What worked was zkao, a purpose-built auditor with heavy context engineering for cryptographic code, and even that needed more than nine and a half hours of scanning.

The bug itself, now tracked as CVE-2026-46669, sat in OpenVM's pairing library for the BLS12-381 and BN254 curves. The verification step failed to check that a scaling factor stayed inside the correct subfield, so a malicious prover could set that factor to an arbitrary value and make a false pairing check pass. In practice that undermines KZG commitments, Groth16 proofs, and BLS signatures, anything built on pairing checks inside an OpenVM guest. Both the AI and the human maintainers rated it critical, and OpenVM shipped a fix in version 1.6.0, with known partners already upgraded.

Why it matters

If you write or depend on zero-knowledge circuits, the lesson is narrow and practical: generic LLM prompting missed a soundness break that a specialized, literature-aware setup caught after hours of work, so treat AI code review as a tool that needs domain scaffolding, not a one-prompt answer.