Will AI Replace Smart Contract Auditors?

By Owen Thurm Published November 6, 2026 5 min read

Quick answer. Not on published evidence. The best agent in EVMbench found 45.6% of known high-severity bugs, and a retest on 22 newer incidents saw no successful end-to-end exploit. AI adds breadth; humans still supply context, judgment and verification.

Watch: Will AI Replace Smart Contract Auditors? 2:00

Video transcript

Will AI replace smart contract auditors? Not on published evidence. AI adds breadth, but humans still supply context, judgment and verification.

In EVMbench, from OpenAI and Paradigm, the best agent found 45.6 percent of known high-severity bugs. The benchmark only credits bugs already reported, and false positives carry no penalty.

Other studies agree. Nethermind's AuditAgent found 30 percent of what its own auditors found. A BlockSec retest on 22 newer incidents detected up to 65 percent, but saw no successful end-to-end exploit.

BlockSec co-founder Yajin Zhou sums it up. AI handles breadth, humans handle depth. Agents scan cheaply and catch common bugs, while humans bring business logic, adversarial reasoning and verification.

Noise is a real cost. On its own EVMbench run, Nethermind's validation cut flagged findings from 3,230 to 798.

Attackers have the same tools. In Anthropic's tests, agents found two new zero-days in 2,849 recently deployed contracts. Cheap scanning works for attackers too.

So the job is changing, not disappearing. Arbitrum offers AI screening before a human audit, Nethermind runs AuditAgent as a pair auditor, and Sherlock blends frontier models with AI-native researchers.

Guardian's Vanguard audit uses both. Team A does manual review and invariant fuzzing. Team B is driven by Helix, Guardian's AI auditor. Humans verify every finding, then a 100,000 dollar Defender contest opens the scope.

Read the full article for the evidence in detail, or talk to Guardian about a Vanguard audit at guardian.security.

AI agents can now find real smart contract bugs, and exploit some of them. So founders, investors and auditors keep asking whether the humans are next.

Paradigm, which co-built the EVMbench benchmark, says “a growing portion of audits in the future will be done by agents.” The open questions are how large a portion, and how soon.

This post answers from published results, not opinion. For the benchmark itself, see EVMbench explained.

The Evidence So Far

Five sets of published results frame the question. Each measures something different.

Study	What it tested	Headline result
EVMbench, OpenAI and Paradigm (Feb 2026)	120 high-severity bugs from 40 audits	Best scores: 45.6% detect, 41.5% patch, 72.2% exploit
EVMbench retest, BlockSec researchers (Mar 2026)	22 real incidents from after mid-February 2026	Up to 65% detected; no end-to-end exploit
Nethermind AuditAgent (Sep 2025)	29 audits, against its own auditors' findings	Found 30% of what the auditors found
Anthropic (Dec 2025)	405 contracts exploited from 2020 to 2025	Working exploits for 207 across 10 models, in simulation
Deakin University (Feb 2026)	Solidity written by three leading models	Severe flaws were frequent, the authors found

If your team writes code with AI, read whether AI-generated smart contracts are safe.

How to read the numbers

Known bugs only. EVMbench's detect mode credits bugs already in the audit report, and false positives carry no penalty.
Possible contamination. OpenZeppelin said top agents had likely seen the benchmark's bug reports in pretraining, Cointelegraph reported.
Not a live-bug rate. EVMbench's authors warn the scores do not mean that share of live bugs is exploitable.
New code is harder. Across 22 newer incidents and 110 agent-incident pairs, the retest found zero successful end-to-end exploits.

What AI Does Better

The studies agree on where agents help.

Breadth. In the words of BlockSec co-founder Yajin Zhou, “AI handles breadth (systematic scanning).”
Cost. In Anthropic's tests, a GPT-5 agent run cost $1.22 on average to scan a contract.
Common bugs. In the retest, six incidents were caught by nearly all agents, The Block reported.
Exploit writing. In Anthropic's tests, 10 models together wrote working exploits for 207 of 405 previously hacked contracts.

Context helps most. Hints that named the broken mechanism sharply raised EVMbench's patch and exploit scores.

Its authors suggest discovery is “the primary bottleneck in many repos.” The retest disagrees: agents detected up to 65% of newer incidents but exploited none.

What Humans Still Do Better

The same studies show where agents fall short.

Business logic. Nethermind says its tool can improve on context-specific bugs, where system logic is critical.
The hard tail. In the retest, four incidents were caught by no agent, and five by only one of eight.
Depth. Zhou's split: “humans handle depth (protocol knowledge, adversarial reasoning).”
Noise. On its own EVMbench run, Nethermind's validation cut flagged findings from 3,230 to 798.

Nethermind puts the last point simply: “Precision is a measure of cost.” Every unverified finding is work for someone.

Humans miss bugs too. In EVMbench's Code4rena data, the number of auditors who found many bugs was “in the single digits.”

So the useful question is not human or AI. It is how many independent reviewers, of either kind, see the code. Our explainer on the hybrid AI and human audit model covers that design.

How the Job Is Changing

On current evidence, the job is changing rather than disappearing. Zhou's summary: “replacing humans is not close.”

Attackers have the same tools

In Anthropic's tests, GPT-5 and Claude Sonnet 4.5 agents found two new zero-days in 2,849 recently deployed contracts. Anthropic's conclusion: “now is the time to adopt AI for defense.”

EVMbench's authors add that better models and tooling “raise the expected attack surface.” Cheap scanning works for attackers too.

AI before and alongside the audit

Arbitrum. Its Security Program, live since September 29, 2026, offers selected teams AI screening before a human audit.
Nethermind. It runs AuditAgent alongside manual audits, as “a pair auditor rather than a standalone solution.”
Sherlock. Its Audit Engine combines frontier models, specialist AI auditors and “AI-native security researchers.”

The skills that gain value

Whichever study is right about the bottleneck, three skills gain value: knowing where to look, verifying what tools report and judging severity.

Those rest on the fundamentals in our guide to becoming a smart contract auditor: building, reading code deeply and knowing the bug classes.

How Guardian Uses Both

Guardian's Vanguard audit puts two independent teams on one scope. Team A uses expert manual review and an invariant fuzzing suite.

Team B is driven by Helix, Guardian's AI auditor. The Vanguard page says it spends “billions of tokens and several days across hundreds of parallel agents.”

Humans verify the context, configuration and depth, and only human-verified findings reach the report. A $100,000 Defender contest then opens the scope to outside researchers and AI auditors.

Guardian has not published a Helix benchmark. For other tools' published evidence, see the best AI tools for smart contract auditing.

Frequently Asked Questions

Can AI audit smart contracts on its own?

Not reliably yet. EVMbench's best agent found 45.6% of known high-severity bugs, and Nethermind's tool found 30% of what its own auditors found across 29 audits. Treat AI output as candidates for a human to verify.

Can AI agents exploit smart contracts?

Yes, in simulation. In Anthropic's tests, 10 models together wrote working exploits for 207 of 405 previously exploited contracts, and EVMbench's top agent exploited 72.2% of its exploit tasks. On 22 newer incidents, a retest found no successful end-to-end exploit.

Should a protocol use an AI audit instead of a human audit?

Not for code that will hold significant value. Run AI scans while you build and before the audit, then have humans verify every finding. Arbitrum's Security Program also offers AI-assisted screening ahead of a full audit.

Is smart contract auditing still a good career?

The work is changing, not vanishing. BlockSec's co-founder says “replacing humans is not close,” while Paradigm expects agents to do a growing share of audits. Nethermind and Sherlock now pair AI tools with human researchers.

Sources

Put AI and human auditors on your code

Vanguard pairs expert manual review and invariant fuzzing with Helix, and only human-verified findings reach the report.

Get a Vanguard quote Explore Vanguard