---
title: "LLM-assisted review"
description: "LLM-assisted review: Emerging tooling. Layer: Circuit / libraries. Characteristic finds: Sustained attention over large cryptographic library surfaces…"
type: "option"
url: "https://zkpick.com/audit/llm-assisted-review/"
section: "03 Auditing a ZK protocol"
authors:
  - "MarketComp"
publisher: "MarketComp"
version: "1.3"
updated: "2026-09-12"
license: "CC-BY-4.0"
json: "https://zkpick.com/data/audit/llm-assisted-review.json"
---

*By MarketComp. Updated 2026-09-12. Version 1.3. CC BY 4.0.*

# LLM-assisted review — *Emerging tooling*

> LLM-assisted review: Emerging tooling. Layer: Circuit / libraries. Characteristic finds: Sustained attention over large cryptographic library surfaces (evidence so far: a single self-reported finding). Blind to: Unknown and unmeasurable; false negatives are invisible. Coverage claim: None. Cost profile: Low per run; expert triage is the real cost. Choose it when: A supplementary sweep alongside manual review and deterministic tooling, especially over large cryptographic library surfaces. Not a substitute for any other category.

Automated review pipelines built on large language models with domain-specific context engineering, run over circuit and cryptographic library code. The category has moved from speculative to evidenced: a missing subfield membership check in an optimised pairing routine was found this way, assigned a CVE, and fixed upstream. *Disclosure: that finding is the work of this document's maintainer, and it is a single self-reported result — weight it accordingly against the peer-reviewed evidence behind the other categories here.*

**Strengths**

- Sustains attention over large codebases where human reviewers fatigue
- Not restricted to a modelled bug class the way SMT and lint tooling is
- Cheap enough to re-run on every change

**Tradeoffs**

- Results depend heavily on pipeline design; naive prompting fails where context-engineered flows succeed
- No soundness or coverage guarantee; false negatives are invisible
- False positives consume expert triage time, which is the scarce resource
- Leading pipelines are proprietary and not independently reproducible, so published claims cannot be checked
- Must not be represented to third parties as equivalent to human or formal review

**Choose it when:** A supplementary sweep alongside manual review and deterministic tooling, especially over large cryptographic library surfaces. Not a substitute for any other category.

Maturity: Emerging; one CVE-assigned production finding · License: Varies; mostly proprietary · Source: https://blog.zksecurity.xyz/posts/openvm-bugs/

**In the what each assurance technique can and cannot find** (https://zkpick.com/audit/#audit-matrix-llm-assisted-review)

| Technique | Layer | Characteristic finds | Blind to | Coverage claim | Cost profile |
| --- | --- | --- | --- | --- | --- |
| LLM-assisted review | Circuit / libraries | Sustained attention over large cryptographic library surfaces (evidence so far: a single self-reported finding) | Unknown and unmeasurable; false negatives are invisible | None | Low per run; expert triage is the real cost |