How We Use AI in Smart Contract Audits, and Where a Human Always Decides
Security Wiz team3 min read
There's one question clients ask us more than most, and they should: do you use AI on our code? The answer is yes, in a limited way, with clear rules. This is our policy, written out plainly.
What AI is good for
An audit includes a lot of repetitive, high-volume work. Scanning tools produce hundreds of warnings that need sorting. Test code has to be written. Confirmed problems have to be written up. That's where AI helps. We use it to:
- Sort scanner warnings, labeling each one as probably real, probably a false alarm, or needing a closer look.
- Draft test code and attack tests, which we then run and read ourselves.
- Draft write-ups of problems a person has already confirmed.
In each case the AI produces a draft or a suggestion. A person reads it.
What AI never does
- It never decides how serious a problem is. Severity is a judgment about how much damage a flaw could do and how likely it is to be used. A person makes that call, using a scoring system anyone can check.
- It never marks a problem as fixed. A finding is closed only after we review and retest your fix ourselves.
- It never sends anything to you without human review. Every finding and every report is read by a person first.
- It never turns a lead into a finding on its own. A warning from a scanner or a suggestion from an AI is a lead, like a tip handed to a detective. It becomes a finding only after a person confirms it.
The reason is accountability. If a report says something is safe or unsafe, a person who can be asked "why?" stands behind it.
What happens to your code
Sending your code to any outside service is a decision, and it should be yours. We call the choice of AI service, and the terms it runs under, the "route."
- Consent first. Your code goes to an AI service only with your written consent, which we collect in the engagement letter you sign before work starts.
- Secrets are stripped. Before anything is sent to any AI service, we remove private keys, wallet recovery phrases, configuration files that hold passwords, API keys, and personal data.
- You can opt out. If you decline, we run the same process on a different route, and the audit is just as thorough.
- Sensitive work gets a stricter route. Code for unreleased products under a strict confidentiality agreement goes through a more restricted route by default.
- It's on the record. The route we used is recorded in the audit log and disclosed in your report.
Why the route matters
AI providers differ. Some store data in particular countries. Some may use what you send them to train their models unless you opt out. How long they keep your data varies, and not all of them say. For code that is already public, that may not matter. For unreleased code under a confidentiality agreement, it can. That's why the route is a choice you make and not one we make for you, and why the engagement letter names it.
Why we use AI at all
Careful review takes human time. The time saved on repetitive work goes back into the parts only a person can do: understanding how your product is meant to work, spotting the flaw in an assumption, and building a working attack to prove it. The aim isn't a cheaper audit. It's to spend human attention where it finds the most.
If you have questions about how your code would be handled, ask before you sign. We'd rather answer them up front. You can also see how our process works in the FAQ, or request an audit.
Shipping something onchain?
Send us your repo and we’ll return a quote within 24 hours.