In a recent AI‑driven security scan, our auditor zkao discovered a critical soundness flaw in Open VM’s guest library openvm‑pairing that allowed a malicious prover to forge any pairing equality. The bug, tracked as CVE‑2026‑46669, was patched in Open VM 1.6.0 and all partners have migrated to the fixed release.

The issue is not in the zkVM’s proving system itself but in the vulnerable library. A quick manual triage confirmed exploitability, and the detailed AI report with a minimal PoC enabled swift disclosure. The flaw occurs because the pairing‑check routine omits a proper subfield check on the scaling factor, letting incorrect pairings pass the verifier.

Open VM’s optimization uses a residue‑witness trick to avoid expensive exponentiation, but the missing check in the optimized equation leaves a gap. The AI‑generated findings Pharaoh and Codex models flagged as Critical or High, yet only the specialized zkao agent could isolate the exploitable bug.

This demonstrates that even sophisticated zkVM codebases can conceal subtle cryptographic errors, which AI tools can uncover when guided by expert‑crafted flows.