Security vulnerability guides for vibe-coded apps

NEXT_PUBLIC_ secret exposure

JWT weak / hardcoded secrets

Verbose error responses leaking stack traces

Dockerfile secrets in ENV / ARG

Secrets interpolated into AI prompt strings

ACCESS CONTROL

7 checks

Who can read or modify your data — and who shouldn't be able to.

Supabase RLS disabled

Supabase service_role key in client-accessible code

Unauthenticated API routes

Mass assignment (req.body spread into DB)

CORS misconfiguration (wildcard / reflected origin)

IDOR — missing ownership checks on resource IDs

GraphQL introspection enabled in production

INJECTION

8 checks

User input reaching code execution, query parsing, or template rendering.

SQL injection

NoSQL injection (MongoDB)

Command injection

Server-side template injection (EJS, Pug, Handlebars)

XSS — stored, reflected, DOM

CRLF injection in HTTP response headers

CSV formula injection

Prototype pollution

SSRF + cloud metadata endpoint access

SERVER-SIDE

6 checks

Server-level vulnerabilities that let attackers reach internal systems or files.

Insecure file upload (no type validation)

Host header injection

TLS certificate verification disabled

Timing-unsafe token comparison

AUTH & SESSIONS

6 checks

Authentication and session management weaknesses that let attackers impersonate users.

Auth endpoints missing rate limiting

OAuth state validation missing (CSRF)

Webhook handler missing signature verification

Insecure password hashing (MD5 / SHA-1)

Math.random() for token generation

Missing security headers (CSP, HSTS, X-Frame-Options)

NODE / FRAMEWORK

6 checks

Node.js and framework-level misconfigurations common in AI-generated scaffolding.

ReDoS — user input passed to RegExp constructor

Code injection via eval() / new Function()

Dependency CVEs (npm audit)

Express trust proxy misconfiguration

GitHub Actions script injection

LLM output passed to eval / exec / SQL

AI / LLM APPS

3 checks

Security risks specific to apps that integrate large language models.

Prompt injection sinks in user-facing features

Missing AI output validation

Missing health check endpoint

OPS & RELIABILITY

5 checks

Configuration gaps that make incidents harder to detect and recover from.

LOW→

Dockerfile missing HEALTHCHECK instruction

LOW→

Destructive SQL migration (DROP / TRUNCATE without backup)

.env files not excluded in .gitignore

Deployment config security (vercel.json, fly.toml, railway)

Lifecycle scripts with shell execution (postinstall, prepare)

SUPPLY CHAIN

6 checks

Indicators that a dependency or workflow has been compromised.

Dependencies from non-registry sources (git:, file:, http:)

Obfuscated code — eval of base64 / hex payload

Binary executables committed to repository

GitHub Actions permission escalation (write-all, id-token: write)

External Actions not pinned to a commit SHA

LOW→

AI AGENT INJECTION

5 checks

Covert instructions hidden in files AI coding agents read — targeting the agent itself.

"Ignore previous instructions" in CLAUDE.md / README / AGENTS.md

Hidden instructions in HTML comments

Unicode direction-override characters hiding text

MCP server config executing shell commands on startup

MCP server connecting to a remote endpoint

Committed git hooks with shell execution (post-checkout, post-merge)

REPOSITORY TRUST

5 checks

Code-execution traps set in a repository that trigger before any review — targeting developers and AI coding agents the moment they clone or open the repo.

Dangerous git config: core.hooksPath, url.insteadOf, filter.smudge

Devcontainer lifecycle hooks with suspicious commands

Setup/install scripts with curl-pipe-to-shell patterns

Submodule URLs pointing to non-standard or suspicious hosts