Production rescue

AI App Security Audit

Short answer

AI-built apps fail security in predictable ways: keys in the browser, missing row-level security, IDOR on APIs, and LLM endpoints that will happily read another tenant’s documents.

What is an AI app security audit?

A review focused on the failure modes of generated apps and LLM features — not a generic ISO checklist. Priority: data isolation and secret handling.

Common AI-app failures

BugWhy generators create itFix
Client-side API keyDemo code copies the quickstartServer route + usage caps
No RLSTables “just work”Policies per tenant
IDORRoutes use IDs from the URLAuthz on every load
Prompt injectionUser text concatenated into system promptIsolation + allowlists

Minimum bar before paying users

No secrets in the client bundle. Every read/write checks the current user (and org). File URLs expire. Stripe webhooks verify signatures. AI routes are rate-limited and logged.

LLM-specific

If users can upload docs, assume they will try to jailbreak retrieval. Do not put other tenants’ chunks in the same unconstrained context. Log token spend per user.

FAQ

Do I need a full pentest before launch?

Not for a 50-user MVP. You do need the items above. Pentest when you have a reason (enterprise buyer, compliance).

Related

Worried you shipped a key in the client?

That’s the #1 finding. We can hunt the rest — or you can start with the checklist on this page.