#142Case study4 min
A security workflow for building with AI
Turn each meaningful AI-generated change into a scoped check, evidence review and human-approved correction.
Hackaiz team · updated 21/11/2026

01Start with what the AI changed
An AI assistant can make a feature work without establishing who should be allowed to use it. Before accepting a change, write down its security boundary: which person may read a record, which role may change it, and which credentials must stay on the server. For example, a working account page does not answer whether one signed-in person can request another person’s records. These are review questions, not claims that your application has a vulnerability.
Use this short note to guide your own testing and the scope of the next Hackaiz analysis. Pay particular attention when the assistant changes sign-in, database access, file handling or a server endpoint. Asking the same assistant to reread its output can help while building, but it is not a substitute for examining the deployed application and its connected components.
02Connect the implementation, not just its address
Open Analyze and identify the application you intend to check. A website and its repository provide different evidence: the deployed site exposes current behaviour, while the repository contains implementation details. A connected database can add information about its access configuration. Connecting one does not mean the others have been examined. Use Connect GitHub for an authorized private repository, or Connect your system for supported code and database connections.
Before any website or API analysis, prove control of the exact domain and give explicit authorization. Add the DNS TXT record named _hackaiz or serve the requested value at /.well-known/hackaiz-verification.txt over HTTPS. You may also ask the AI editing your project to create that file, then publish the change. A repository connection does not replace domain proof.
03Choose checks, then read their actual coverage
On Analyze, review the selected check families before starting. Clear a family only when you intentionally want to leave that subject outside this run. Unticked checks are not run and are not billed. Analyses and other operations use credits; review the displayed selection and authorization before committing to the action. Do not treat a smaller selection as a complete assessment of the application.
When results arrive, read coverage alongside findings. An incomplete check is not evidence that a control works. Review the supporting evidence and confidence before sending a finding back to your coding assistant. Prioritize Danger, then To fix, while keeping To watch items visible. Preserve the distinction between a demonstrated observation and a question that still needs investigation.
04Keep a person in charge of corrections
In Corrections, examine the proposed change against the original boundary you wrote down. For an eligible code correction, Hackaiz prepares a file preview. Human approval authorizes publication as a draft GitHub pull request, a proposed change awaiting review. It is never merged automatically. Review the affected files and have your developer check that the intended user journey still works.
Test both allowed and denied actions before your team merges and deploys. If the issue concerns a database setting or another external configuration, do not assume a code proposal changes that service: Hackaiz does not write directly to customer databases. Apply the appropriate configuration change through your own controlled process and verify the deployed result rather than only the proposed code.
05Repeat after meaningful changes
After deployment, recheck the affected scope to find out whether the original evidence still holds. A proposal, a merged branch and a deployed fix are different milestones. Keep your review note with the relevant result so the next AI-generated change can be compared with a concrete security expectation rather than a vague instruction to “make it secure”.
Monitoring is optional, uses credits and is limited to connected systems with current authorization and any required domain proof. It can support follow-up as those systems change; it does not remove the need to review releases. This workflow organizes decisions around observed evidence. It does not promise exhaustive coverage or certify the application.
Frequently asked questions
Take the next step
Was this article helpful?