#055Learning3 min
Security mistakes AI can introduce into an app
Review generated access rules, exposed secrets and dependencies before trusting a working prototype.
Hackaiz team · updated 21/11/2026

01A working feature can still expose data
An AI-built application can pass a demonstration while enforcing the wrong access boundary. A page that hides another customer’s invoices is not enough if the underlying request still returns them. Treat generated code as a proposal to review, especially when a change adds accounts, administration, file uploads or database access. The important question is not only whether the screen works, but which identities can reach the data behind it.
Hackaiz can examine authorized GitHub code and supported connected systems for evidence of security mistakes. It does not decide that a vulnerability was caused by AI simply because an AI editor was used. Use the finding’s code location or recorded response to understand the actual defect. A generated application and a handwritten application need the same care around secrets and permissions.
02Look first at access rules and privileged keys
Supabase examples include a table with Row Level Security, or RLS, turned off, a policy that lets every signed-in account read everybody’s rows, and a service-role key bundled into browser code. RLS means rules that decide which database rows an identity may use. The service-role key is privileged: it must not become a shortcut for making a browser feature work. A public key and a privileged server key are not interchangeable.
Also review profiles that let users change their own administrative role and public storage containing private documents. These are concrete access-design errors, not cosmetic warnings. When you review an AI-generated fix, explain the intended rule in ordinary language first: for example, a customer may read their own invoices, while support access follows a separate authorized path. That intention is needed to judge whether returned data is actually exposed by mistake.
03Review the change before adding more code
Open Analyze to prepare an application assessment, or use Connect GitHub for the private repository workflow. Website and API analysis requires proof that you control the domain: a DNS TXT record at _hackaiz, an HTTPS file at /.well-known/hackaiz-verification.txt, or an instruction to your project’s AI to create that file. Publishing a URL or asking an assistant to review it is not ownership proof.
If the generated change adds a Skill, an MCP configuration or a dependency manifest, use Check before installing for a separate static component review. It reads submitted text without running the component. This is useful for hidden commands, secret reads and suspicious installation instructions, but it does not replace analysis of the app’s own access rules.
- 1Identify which data and permissions the generated feature changes.
- 2Connect the relevant repository or authorized system and complete any required domain proof.
- 3Choose the checks you want before launching; unticked checks are neither run nor billed.
- 4Read the evidence alongside the changed code rather than asking the AI to silence the warning.
04Keep control of the repair
Application report findings use Danger, To fix and To watch. Start with the evidence and affected access path, not just the label. A rejected anonymous request proves that particular request was refused; it does not establish that every account, table or endpoint is protected. Keep any untested path on your review list rather than treating the absence of a finding as assurance.
When Hackaiz proposes GitHub fixes, they are draft pull requests. A human reviews and approves them, and they are never merged automatically. Check that the proposed change still allows the intended customer workflow and does not simply remove a necessary permission indiscriminately. Checks consume credits; repeating an assessment is a new paid action. Optional monitoring applies only to connected systems and should not be confused with a firewall that blocks attacks.
Frequently asked questions
Take the next step
Was this article helpful?