AI-assisted security audit of the Zcash ecosystem
At ZCG's request, Least Authority audited the Zebra, zcashd, librustzcash, orchard and halo2 repositories with an AI-assisted approach. The scope, methodology and report are summarized.
At the request of Zcash Community Grants (ZCG), Least Authority carried out an AI-assisted security audit of security-critical repositories in the Zcash ecosystem and published the methodology and results. The aim was to explore how much AI workflows can support large-scale security review, on the premise that humans verify carefully. The five target repositories are zebra, zcash, librustzcash, orchard and halo2, with focus on consensus safety, wallet key material, denial of service and proof system soundness.
The methodology has AI agents explore the repository with structured prompts, record potential issues (leads) in a shared table, and put them through human-led triage and verification. It started with a broad sweep of the whole repository, then was gradually refined by adding per-project threat models based on documentation, GitHub issues and CVEs, and focusing on specific attack scenarios and selected code areas. Each lead went through duplicate checking, attempts at verification using existing and newly created adversarial test vectors or PoCs, manual review and OWASP-based impact and likelihood assessment, and the final judgment was made by experienced auditors.
Least Authority also developed a custom toolkit for AI audit workflows. It provides a standard workspace holding scope, threat model, run logs, lead tracking and detailed findings, works with various models such as Claude Code and Codex, and considered sandboxing because agents run commands freely. The toolkit repository is private and shared only with some developers, with plans to release it later. Of the seven weeks, two were spent on toolkit development, two to three on lead collection with Claude Opus 4.7 and OpenAI GPT 5.4, and the last three on triage and report writing.
The final report was delivered on May 29, 2026. Confirmed issues are 17 in zebra (5 High, 8 Medium, 1 Low, 3 Informational), 11 in orchard, 10 in zcash, 7 in librustzcash and 2 in halo2 (1 High, 1 Low). Two issues in the interim report were already being handled by the Zcash team as private security advisories. Many duplicates and false positives were filtered out and there is room for improvement. It concludes that the value lies in the combination of a structured workflow, continuous improvement as models advance, and human security expertise, not in automation itself.
This summary is based on the original; check the original for exact details.