Is your smart contract audit an AI scan? A 20-minute test

Table of Contents

A garage can plug a code reader into your car, print what it says and hand you the sheet. The reader is a good tool. The sheet is not a service, and nobody has looked at your brakes.

Some smart contract audit reports are that sheet. A tool or a language model ran over the code, its output went onto a letterhead, and the document was sold as an audit. You can usually tell from three places: the published report, the quote and the first sales call. This guide covers what to look for in each, plus a 20-minute test you can run on your own code before you pay anyone.

A disclosure, because it should change how you read this. Fidesium sells manual smart contract audits, so we sit on one side of this question. Every sign below is one you can check yourself, on anyone’s report, ours included.

Published 3 November 2026. Last updated 3 November 2026.

Quick Answer

A smart contract audit is people reading your code against what the protocol is meant to do. A scan is a tool or a language model matching known patterns. A report is probably a scan if its findings match what a free tool prints on your code, it names no commit or reviewers, and nothing in it concerns your protocol’s own logic.

What separates an AI scan from a smart contract audit?

A scan checks code against patterns someone has already written down. An audit asks whether the code does what the protocol is meant to do, and that question has no pattern. Both are useful. Only one of them finds the bugs that drain protocols.

The clearest public number on the gap comes from Zhuo Zhang and colleagues, who studied 516 exploitable bugs from 167 real contracts reported or exploited in 2021 and 2022. Their finding, presented at ICSE 2023: “More than 80% exploitable bugs are beyond existing tools.” They call these machine unauditable bugs. Think of a price that can be moved inside one block, accounting that drifts by a rounding direction, or a check present in one function and missing from its twin. A detector has no rule for any of them, because the rule lives in your design.

Automated and AI-assisted review is not the problem, and firms that use it well say so. Sherlock’s docs, for one, offer “a fast AI-only pass in as little as 2 days” as its own option, next to an engagement that includes human security researchers. That is honest labelling. What goes wrong is the same kind of output sold under the word “audit”, with nothing on the page to tell you a person never read your code.

For how a full audit runs, step by step, see how a smart contract audit works. This page is only about telling the two apart.

Before you read on, open one of our published audit reports in another tab. Everything below is easier to see with a human-read report next to the one you were sent.

What does a scan look like inside a published report?

A scan gives itself away in the scope, the finding titles and the fixes. Start with the firm’s published reports, not its landing page. Pick two reports for different protocols and read them side by side.

Where to lookWhat a scan tends to showWhat a human review tends to show
ScopeNo commit hash, or “the contracts provided”A repository, a commit hash and a file list
Finding titlesDetector names reworded: reentrancy, floating pragma, missing zero-address check, variable could be immutableTitles about your protocol, such as “Liquidations can be skipped when the oracle is stale”
Severity mixMostly Low, Informational and gas notes, with no serious finding about logicSerious findings about how value moves through the protocol
DescriptionsA paragraph explaining the bug class in generalYour function names, a call path, often a proof of concept
FixesThe same advice in every report (“add a reentrancy guard”, “check for the zero address”)A fix for this code, then a fix review with a status per finding
System overviewMissing, or one sentence lifted from your READMEHow the protocol works, who can call what, and what it assumes
Method and people“Our proprietary engine”Named methods and named reviewers
Two reports side by sideSame structure, same sentences, different protocol namesDifferent shapes, because the code was different

No single row proves anything. A real audit carries low-severity findings too, and a good report may include tool output as well. The pattern is what counts. When every row lands in the middle column, nobody has read the code.

Findings that cannot exist in your code

Language models add a tell of their own: findings that are confident, well written and wrong. Three are quick to check.

  • Overflow on Solidity 0.8 or later: Since 0.8.0, “all arithmetic operations revert on over- and underflow by default”. An overflow finding on that code is only valid inside an unchecked block. A report that flags overflow on plain arithmetic, or recommends SafeMath, was written for older code.
  • Functions and lines that are not there: Search your repository for every function name and line number the report cites. A model will sometimes describe the code it expected rather than the code you wrote.
  • A real bug from somewhere else: The right vulnerability class with the wrong contract, or details that match a famous incident better than your code.

The first check takes a minute. In this contract, only one line can overflow:

// SPDX-License-Identifier: MIT
pragma solidity ^0.8.20;

contract Vault {
    mapping(address => uint256) public balance;

    function deposit() external payable {
        // Checked by default since 0.8.0: an overflow here reverts.
        balance[msg.sender] += msg.value;
    }

    function _credit(address user, uint256 amount) internal {
        unchecked {
            // Wraps silently. An overflow finding belongs here, nowhere else.
            balance[user] += amount;
        }
    }
}

Security programmes have seen this kind of output at volume. Immunefi, the bug bounty platform, reported in July 2023 that “not a single real vulnerability has been discovered through a ChatGPT-generated bug report submitted via Immunefi”, and it permanently bans accounts that submit them. Daniel Stenberg, who maintains curl, said one AI-written report “reeks of typical AI style hallucinations” and wrote of another, “There was no buffer overflow.” Checking every cited function against your own code still takes ten minutes, and it is the cheapest check on this page.

Run the free tool before you read the paid report

If every finding in the paid report has a twin in the free output, and nothing in it is about what your protocol is meant to do, you paid for a scan. It may be a good scan. It is still not an audit, and it should not cost what one does. Here is the test.

  1. Install Slither, the open-source static analyser for Solidity from Trail of Bits. ethereum.org has a step-by-step Slither tutorial.
  2. Run it on the same commit the firm reviewed, or the one you are about to send.
  3. Note each detector name in the output. Slither’s own detector documentation grades them: reentrancy-eth is High, missing-zero-check is Low, solc-version and naming-convention are Informational, and immutable-states is an optimisation.
  4. Paste the same contracts into a public chat model, ask it to audit them, and save the answer.
  5. Read the paid report finding by finding, and mark each one that has a twin in either output.

For Rust on Solana, use whichever open-source linter your team already runs. Any tool will do, as long as you compare its output with the report. For what each free scanner is good at, see our comparison of Solidity static analysis tools.

Here is how that plays out, as an illustration rather than a client story. A team with a lending market of about 2,000 lines of Solidity has a launch in three weeks. They are quoted an audit delivered in four days.

The sample report lists 31 findings: 24 Informational, 6 Low and 1 Medium, a reentrancy note on a function that already updates state before it calls out. Their lead dev runs Slither on the same commit over lunch.

Twenty-eight of the 31 titles have a twin in the output. Not one finding mentions liquidations, the oracle or the interest model, which are the three places a lending market loses money. They have their answer before the call is booked.

What does the quote tell you?

A quote describes the work before anyone has done it, so it is the cheapest place to catch a scan. Four signs:

  • The price does not move with the code: A flat fee per contract, whatever its size, prices a tool’s run time. Human review is sold in reviewer time, and that follows the line count and the complexity.
  • Delivery does not move with the code either: A few days for a few hundred lines is plausible. A few days for tens of thousands is not. Our Adrena audit covered 33,747 lines of Rust and took five weeks of assessment, followed by a separate review of the fixes.
  • Nobody asked you anything: Real scoping asks for the commit, the file list, the documentation and what the protocol must never allow. A quote that arrives before anyone asked what your protocol does was written without needing to know.
  • A badge instead of a fix review: Remediation is where a report earns its price. A quote that includes a certificate or a badge but no review of your fixes is selling the logo. Ours includes two rounds of fix verification.

For the contract terms to settle before you sign (the commit, the reviewers, the fix rounds and the re-review policy), our shortlist of smart contract audit firms sets out four questions. Pinning the commit also matters after delivery, because an audit expires when the code or its assumptions change.

Want a quote you can hold against the others? Send the repository, the commit and the file list through our manual audit page, and we will come back with a scope and a price.

Five questions that expose a scan on a sales call

Ask what they would look at first in your protocol. A reviewer who has not seen your code will still have a view on where a protocol like yours breaks. A sales script will not. The other four questions back it up.

QuestionA reviewer’s answer sounds likeA scan’s answer sounds like
Who will read our code, and what have they shipped in our language?Names, and reports you can open“Our team of experts”, or the name of an engine
What would you look at first in our protocol?Something specific: the oracle, the liquidation path, the upgrade keyA list of vulnerability classes that fits any protocol
Can you walk me through how one finding in a past report was found?What looked wrong, what was tested, how it was confirmed“Our engine flagged it”
What do you need from us before you start?Docs, a commit, a file list, invariants, a contact for questionsA repository link
What happens if you find nothing serious?An honest account of what was covered and what was not“Then you are safe”, or a badge

What a human-read report shows that a tool cannot

A human-read report names the commit, the methods and findings about what the code means. Our reports are public, so here are two to hold next to the one you were sent.

Adrena, a Solana perpetuals exchange

The report page names the commit it covers, the size of the code (33,747 lines of Rust, 168 files, 125 public instructions) and the dates: 10 February to 18 March 2026, then a review of the fixes from 31 March to 8 April.

It lists the nine methods applied, from manual source review and economic modelling to formal verification. It records 42 findings and a risk score that fell from 43 out of 100 to 11 as the fixes were verified.

One finding: the program overwrote an oracle price’s timestamp at the moment it stored it. With a freshness setting the caller could choose, that opened a window of roughly 20 seconds in which stale prices passed the check. A scanner has no rule for it, because the code ran exactly as written.

Yimmit, a custodial exchange

In this penetration test, the most serious finding was a public Fireblocks deposit webhook that credited balances “from amounts and addresses taken straight from the request body”, with no signature check. An attacker could mint balances with no deposit and then withdraw real assets. We disclosed it to Yimmit in an urgent notice on 5 June 2026, ahead of the report. No detector has a rule for “this endpoint trusts whoever calls it”, because whether it should is a design question.

When you hire us, you get the core team, not a contractor running a scan. The people who built our tooling read your code.

Our automated layer is deterministic static and abstract syntax tree (AST) analysis, so the same commit gives the same findings twice, and it keeps running on every commit after the report. It clears the mechanical ground. The audit is the reading.

Frequently asked questions

Is an AI smart contract audit a real audit?

Not on its own. An AI smart contract audit is a scan: it finds known patterns quickly and cheaply, and an honest firm labels it that way. It becomes a problem when it is sold as an audit, because one study of 516 exploitable bugs found more than 80% beyond existing tools.

Can automated tools find every smart contract vulnerability?

No. Tools match known patterns such as reentrancy or unchecked calls. Zhuo Zhang and colleagues studied 516 exploitable bugs from 2021 and 2022 and found more than 80% were beyond existing tools. Our reading: a tool cannot know what the protocol is meant to do.

How can I tell if a smart contract audit report is genuine?

Look for a commit hash and file list, named reviewers and methods, findings about your protocol’s own logic, and a fix review with a status per finding. Then run a free scanner on the same commit. If every finding in the report has a twin in the free output, it is a scan.

Is it bad if an auditor uses automated tools?

No. Automated analysis is a normal early step in an audit, used to clear the known patterns, and good reports say which tools ran. The question is what came after: whether people read the code against its design, and whether the report shows that they did.

How long should a smart contract audit take?

Scope decides it. A few hundred lines can take days. Our Adrena audit covered 33,747 lines of Rust and took five weeks of assessment, then a separate fix review. A firm that quotes the same few days for any codebase is pricing a tool’s run, not reviewer time.

What should a smart contract audit report include?

The commit and files in scope, a system overview, the methods used, and for each finding a severity, a code-specific description and a fix. Then the status of every finding after the fix review. A report without a commit hash cannot tell you which code it describes.

Get a quote

Send us the repository, the commit and the file list, and we will come back with a scope and a price. Then hold our published reports next to anyone else’s and judge the difference yourself.

Request a quote · Read our published audit reports

Share:

More Posts

Tell us your security needs