The best AI smart contract auditors, compared on public evidence

Each tool's benchmark is quoted exactly as its maker publishes it, with the denominator they use. The EVMBench paper describes 117 curated vulnerabilities from 40 repositories; Nethermind reports against the earlier set of 120. Percentages on different denominators are close but not identical, so read the counts.

Tools are ordered by published EVMBench Detect score, then alphabetically for tools that publish none. Pricing, coverage and delivery come from each vendor's own site. Where a vendor states a figure without naming the test behind it, the table says so.

Higher numbers exist that measure something else. SolidityGuard, an open-source project, reports 120 of 120 on EVMBench as ground-truth coverage by its pattern scanner. That is a different measurement from the graded audit report the Detect task scores, so it is not ranked alongside the others. Frontier models used on their own score lower: the EVMBench paper reports no base model above 46% on Detect.

Cecuro coordinates about 180 specialized AI agents over an average of 8 hours per audit. They investigate the code from adversarial angles and validate each finding before it reaches the report, which is severity-classified and signed. On EVMBench, Cecuro reports 91.45% detection, measured in its own run on the same dataset and evaluation criteria as the published model baselines. On a separate set of 90 contracts exploited in real post-2024 attacks, $228M in losses, it flagged the exploited vulnerability in 83 (92%), covering $96.8M in exploitable value against $7.5M for a baseline agent. Pricing is published on the pricing page, and the method is on the benchmarks page.

Azimuth is TestMachine's offensive security agent. It maps a codebase's attack surface, generates exploit hypotheses and runs them against forked mainnet state, and the product page says confirmed findings include an attack path, an execution trace and a reproduction. TestMachine reports 86.3% on EVMBench Detect (101 of 117) from a run on 14 July 2026, and it has published sanitized logs of that run on GitHub, which makes it the most inspectable EVMBench result on this page. It covers Solidity and Vyper on EVM chains and Rust on Solana, and offers a free daily quota.

AuditAgent is Nethermind's AI security analysis tool, which its own docs describe as a pair auditor and not a replacement for a skilled human reviewer. Nethermind ran it on all 40 EVMBench repositories and reports 80 of 120 vulnerabilities found, 67% recall after its validation phase; before validation recall was 74%, with about four times as many findings to read. It supports Solidity on EVM chains, Cairo on Starknet and Rust on Solana, and publishes scan pricing of $0.05 per BLoC.

Almanax calls itself an AI security engineer. It installs as a GitHub app, scans existing code and reviews each pull request, can block risky code before merge, and adds threat modeling and custom rules written in plain English. Its site lists ecosystem-specialized models for EVM, Solana, Stellar and Aptos and offers a 7-day free trial without a credit card. It publishes no benchmark figures. Its homepage also states that Almanax is joining depthfirst, so check the product's status before you commit.

Octane describes itself as an AI-powered offensive security platform. You connect a repository, it builds a model of how the system runs, and it analyzes every pull request in context, with findings that include the vulnerable path, root cause, severity and a suggested fix. It integrates with GitHub, GitLab and CI/CD, lists ten languages from Solidity and Rust to Go and C++, and covers EVM chains, Solana, Aptos, Sui and Cosmos, plus bridges, validator software and node clients. It states a 97.4% true positive rate without naming the benchmark behind it. Pricing depends on codebase size, cadence and integrations, and is not published.

Olympix positions itself as pre-deployment security tooling that analyzes smart contracts deterministically, with static analysis, generated unit tests and mutation testing. Its BugPoCer tool combines custom detectors, symbolic execution, fuzzing and AI. Olympix's own October 2026 report says BugPoCer found 94 of 160 published audit findings across fifteen protocols, against 53 for Claude Opus 5.5 used on its own; that is a vendor-run test on its own dataset, not EVMBench. Its materials reference Solidity only, and it offers a free static analyzer.

Sherlock AI is the development-stage product from Sherlock, the audit contest platform. It connects through a GitHub app and scans every commit and pull request; Sherlock says it is trained on thousands of real vulnerabilities and findings, and its launch post describes it as an open beta. Sherlock also offers Audit Engine, announced in August 2026, which combines AI auditors, advanced models and researchers in one review and consolidates their findings into one result. Sherlock publishes no benchmark score or pricing for either.

V12 was introduced by Zellic in September 2025 as an autonomous Solidity auditor and now runs at v12.sh as an autonomous security system. It reviews every pull request through a GitHub app and also ships a CLI, a web app and an MCP integration for coding agents. New accounts start with $100 in free credits and runs are priced individually; the site shows one example run at $18.40 for a 2,841-line repository. V12's own Crypto Security Bench, run in June 2026, reports 181 of 289 known issues found (62.6%) across Rust, C++, Solidity and Go, against 36.0% for a Claude Code harness. That is a vendor-run benchmark on its own dataset, not EVMBench.

Treat any AI auditor as a fast, repeatable first layer. Benchmarks like EVMBench measure detection of known vulnerability classes, not judgment about a mechanism nobody has deployed before, and a clean report from any tool on this page is evidence, not proof that no bug remains.

For novel mechanism design, formal verification, or a launch where investors require a named firm, add a manual review or an audit contest. AI vs manual smart contract audits covers when that is worth it, and the 2026 smart contract audit cost guide covers what it costs.