ONLINEAGENT_OPS 2026.Q3 HOME HUBSMODELSBUILDINGCOMPANIESFILMMAKINGSECURITYLIVING WITH ITSKILLSALL HUBS, ONE TERMINAL → ARTICLES CRAFT RECORD BLOG TOOLS SKILLS SEARCH /
HOME›ARTICLES›AI APP SECURITY CHECKLIST: A PRINTABLE ONE-PAGE
ARTICLES · CHEAT SHEETINTERMEDIATE

AI App Security Checklist: A Printable One-Page Version

A printable one-page security checklist for apps built with AI: secrets, logins, input checks, prompt injection, permissions, packages, data and logs.

READ3 min
WORDS589
SECTIONS5
TYPECHEAT SHEET
CHECKED13 OCT 26

The security checklist, on one page. A security checklist for AI-built software is already a checklist; this is its printable one-page version: the same three tiers in the same order, shortened, plus the agent rules from prompt injection and permission tiers. Print it (Ctrl+P, or Cmd+P on a Mac) and tick it before you launch.

TL;DR — THE SHORT VERSION
  • Check in order of what a mistake costs. Tier 1 before it goes public, tier 2 before other people’s data, tier 3 before money or logins. Stopping early is fine; skipping ahead is not.
  • Anything the AI reads or writes is a stranger’s text until you have checked it.
  • Scanners help, but a person who reads code still has to read it.

Tier 1 · Before it goes public

  • ☐ Secrets: search the whole project for API keys, tokens and passwords. AI-written code often puts them inline.
  • ☐ History: a key deleted from a file is still in version history. Replace the key.
  • ☐ Packages: each one exists, is maintained, and has the name you expected. Models invent package names, and attackers can register them. 1
  • ☐ Debug mode off; no error details shown to users.
  • ☐ CORS not wide open (the setting for which other sites may call your app).

Tier 2 · Before other people’s data

  • ☐ Input validation on the server. Checks in the browser are a convenience, not a control.
  • ☐ Database queries parameterised, never built by joining strings.
  • ☐ Output escaped where it is shown.
  • ☐ Login checked on every protected page, not only on the page that links to it.
  • ☐ Authorisation checked too: logged in is not the same as allowed.
  • ☐ Rate limits on anything that sends, writes or costs money.
  • ☐ Short error messages: “Invalid credentials”, not “no user with that email”.

Tier 3 · Before money, logins or personal records

  • ☐ Logs: nothing sensitive is logged.
  • ☐ Secrets come from the environment, and the environment is not in the repository.
  • ☐ A person approves anything that cannot be undone.
  • ☐ Restore tested, not just the backup made.
  • ☐ Someone who reads code has read it.

The AI-specific checks

  • ☐ Prompt injection: text your app reads (pages, emails, files, comments) can carry instructions the model cannot reliably tell from yours. It is first on OWASP’s list of AI risks. 2 Keep fetched text out of the system prompt and mark it as material, not orders.
  • ☐ Permissions: no complete fix exists, so OWASP’s own measures are to limit the damage: fewest privileges, and a person’s approval for risky actions. 3 Set access by how hard an action is to undo, not by how much you trust the AI.
  • ☐ Skills and plug-ins: read each one before installing. A February 2026 audit found 13.4% of published agent skills had a critical security issue. 4
  • ☐ Model output: treat it like text from a stranger before it reaches a command line, a database query, a file path or a browser.

Know the limits. This is not a penetration test or legal advice. For payments, health data or anything regulated, start here; do not stop here.

TAKEAWAY

Ask of every feature: what does it read that a stranger can write, and what could it do if it obeyed? Limit the second.

SOURCES
  1. 1Spracklen et al., We Have a Package for You! A Comprehensive Analysis of Package Hallucinations by Code Generating LLMs, USENIX Security 2025, read at source 23 Sep 2026 (as cited on this site’s vibe coding page): “the average percentage of hallucinated packages is at least 5.2% for commercial models and 21.7% for open-source models”; the paper calls this “a novel form of package confusion attack”.↑
  2. 2OWASP, LLM01: Prompt Injection, GenAI / LLM Top 10, 2026 edition, read at source 11 Sep 2026 (as cited on this site’s AI agents for beginners page): prompt injection is LLM01, first on the list.↑
  3. 3OWASP GenAI Security Project, LLM01:2025 Prompt Injection, undated page, read at source 12 Oct 2026 (as cited on this site’s prompt injection meaning page): “it is unclear if there are fool-proof methods of prevention for prompt injection.” “Enforce privilege control and least privilege access”; “Require human approval for high-risk actions”↑
  4. 4Snyk, “ToxicSkills”, 5 Feb 2026, quote re-read 22 Sep 2026 (as cited on this site’s AI agents for beginners page): “13.4% of all skills, or 534 in total, all contain at least one critical-level security issue”.↑

Spracklen et al., We Have a Package for You!, USENIX Security 2025, read at source 23 Sep 2026.

The three tiers and their order are this site’s own method, from the security checklist page; the order is argued from what each mistake costs, not measured.

Free printable AI cheat sheets, and AI news every week.

Sixteen one-page AI guides and checklists in one free PDF, no sign-up. And if you want to keep up, the newsletter covers what changed in AI each week, in plain English, with the sources.

SUBSCRIBE FREE ↗ GET THE CHEAT SHEET PACK
aipromptgeneer://$ about this siteAI PROMPT GENEERAI explained in plain English. Every claim links to its source, with the date it was checked.NEWSLETTERWhat changed in AI each week, in plain English.SUBSCRIBE FREE ↗AI PROMPT GENEER · MACHINE ASSISTANCE DISCLOSEDEVERY ARTICLE CARRIES ITS OWN CHECKED DATE