00AI vs manual audits

An AI audit is not a cheap manual audit.

It is a different instrument, good at a different half of the problem. Guardix is repeatable, comes back in hours, and takes a position on every finding it raises. A person still leads where the attack is bespoke to the way you built this. Here is exactly where that line falls.

$200 flat per audit · results in 1–3 hours · re-runs priced the same

01Where the line falls

A machine can settle a bug that has a shape.

Reentrancy has a shape, and it is the same shape in your code as in the last hundred exploits. The attack that only works because of how you designed this has no shape until somebody invents it.

This half is swept on every run.

Reentrancy
An external call made before internal state is updated — across functions and across contracts.
Oracle and price manipulation
Staleness windows, single-source spot reads, and flash-loan-inflated prices feeding liquidation math.
Access control
Unprotected entrypoints, missing timelocks, and role boundaries an attacker can step across.
Accounting drift
Internal accounting diverging from the balances the contract actually holds.

Spend an auditor's hours on this half.

Novel, bespoke attack paths
A sequence that only works against your protocol. There is no prior exploit to match it against.
Cross-protocol composition
Assumptions that hold in isolation and break when your contract is one leg of somebody else's stack.
Economic and game-theoretic design
The code is correct and the mechanism is still exploitable. That argument happens with your team, not against your repo.
Intent and spec gaps
The contract does exactly what the spec says, and the spec is wrong. Someone has to ask you what you meant.
Guardix will not find the attack that exists only because of the way you built this. Nothing that pattern-matches will. That is the thinking you are hiring a person for.
02What only one side publishes

One of these audits shows you what it decided not to report.

Below is the complete verdict ledger of one real Guardix run — 138 findings on a lending protocol, counted out of the report you can open. 19 of them are findings Guardix raised and then argued itself out of.

Verdicts by severity across all 138 findings in the Guardix sample run
Severity Confirmed Dismissed Design tradeoff
critical 1 1 0 0
high 48 34 2 12
medium 65 28 10 27
low 23 5 7 11
info 1 0 0 1
All findings 138 68 19 51
GDX-69 · high

Yield repay tops up shortfall from nftOwner without balance-delta guard before pool yieldRepayERC20 pull

YieldStakingBase

GDX-91 · high

External interest rate model is a single point of failure per asset group

InterestLogic

Both are high severity. Neither is in the confirmed column. A manual report is the findings that survived — the leads a person chased and dropped stay in their notes, and that is entirely reasonable. It just means only one of these two hands you the whole ledger, and a claim you can check beats a claim you cannot.

Open the full report — all 138 findings →
03Side by side

The tradeoffs, row by row.

Two rows go to the manual column, and they are the two that decide whether an engagement was worth booking. A table that hands every row to one side is an advert, not a comparison.

Guardix compared with a manual smart contract audit, dimension by dimension
Turnaround 1–3 hours, on demand 2–4 weeks, booked in advance
Price per run $200 flat, any repo size $50,000+ per engagement
Running it again Any commit, the same flat $200, diffed against your last scan A new engagement
Mechanical bug classes Reliable, and swept again on every run Reliable
A stated verdict on every finding Confirmed, dismissed, or design tradeoff — all 138 of them in the run above The findings that survived, written up
Novel, protocol-specific attack paths Weak. There is no prior shape to match against This is the thing you are hiring them for
Economic and governance design Raised, and usually filed as a design tradeoff rather than a bug Argued through with your team
What you walk away with A versioned report anyone can open — a verdict on every finding, and a file and line range on 127 of the 138 A PDF, pinned to one commit
  1. Run Guardix before you book the engagement.

    Everything with a stable shape is cleared and already carries a verdict, so the expensive weeks don't open with reentrancy.

  2. Re-run it on the commits that matter.

    The same flat $200 each time, back in 1–3 hours, diffed against your last scan — so the write-up doesn't go stale the moment you ship the fix.

  3. Buy the auditor's hours for the two rows this table hands them.

    Novel attack paths, and economic and governance design. That half is judgment, and no model should be trusted with it alone.

Clear the half with a shape. Pay a person for the other half.

A full audit in 1–3 hours for $200 flat, with a verdict on every finding — so the human engagement starts on the questions only a human can answer.