"Vibe coding" — describing what you want in natural language and letting an AI assistant generate the app — has made software creation astonishingly fast. It has also produced a wave of applications shipped to production by people who cannot fully read the code they deployed. When speed outruns understanding, the result is not just messy code; it is a genuinely dangerous attack surface. Here is what tends to go wrong, and why it matters.
The core problem: code you don't understand, you can't secure
AI assistants are trained to produce code that works, not code that is safe. They happily generate plausible, runnable solutions that omit authentication, trust user input, or hardcode a secret — because the prompt asked for a feature, not a threat model. If no one on the team can review what was generated, those flaws ship silently.
The most common dangers
1. Exposed secrets and API keys
Generated code routinely embeds API keys, database passwords, and tokens directly in source — then gets pushed to a public repo or shipped in a client-side bundle. Leaked keys are scraped within minutes and turn into cloud bills, data theft, and account takeover.
2. Broken or missing authentication and access control
Broken access control is the most prevalent web vulnerability class, and vibe-coded apps are full of it: endpoints with no auth check, client-side-only "security," users able to read or modify other users' records by changing an ID. The app looks fine in a demo and is wide open in production.
3. Injection — SQL, command, and prompt
String-concatenated queries invite SQL injection; unsanitized shell calls invite command injection; and AI features that paste untrusted text into a model invite prompt injection. Generated code often takes the shortest path, which is usually the unsafe one.
4. Cross-site scripting and unsafe rendering
Rendering user content without escaping, dropping dangerouslySetInnerHTML in to "make it work," or trusting third-party data leads to XSS that hijacks sessions and steals data.
5. Exposed databases and over-permissioned cloud
Publicly readable storage buckets, databases with default credentials, and cloud roles granted blanket admin rights are classic AI-suggested shortcuts. One misconfigured permission can expose an entire dataset.
6. No input validation or error handling
Without validation, malformed input crashes the app or corrupts data; without proper error handling, stack traces leak internal details that help attackers map the system.
7. Vulnerable and hallucinated dependencies
Assistants pull in outdated, abandoned, or occasionally non-existent packages — a gap attackers exploit through "slopsquatting," registering the hallucinated package name with malware inside.
8. No tests, no logging, no monitoring
Speed-first builds skip automated tests and observability. When something breaks or is breached, there is no signal, no audit trail, and no way to tell what happened.
9. Privacy and compliance gaps
Collecting personal data with no consent flow, no retention policy, and no encryption is easy to generate and hard to defend — and it runs straight into GDPR, the EU AI Act, and sector rules.
10. Invisible technical debt
Even when it runs, unreviewed generated code is often duplicated, inconsistent, and architecturally fragile — expensive to maintain and risky to change, because no one knows what depends on what.
How to vibe-code responsibly
- Never commit secrets — use environment variables and a secrets manager, and scan every commit.
- Require human code review for anything that touches auth, data, or money.
- Run automated security scanning (SAST, dependency, and secret scanning) in CI.
- Enforce authentication and server-side authorization on every endpoint.
- Validate and sanitize all input; escape all output.
- Apply least privilege to databases, cloud roles, and any AI agent's tools.
- Pin and verify dependencies; reject packages you cannot confirm exist.
- Add tests, logging, and monitoring before launch, not after an incident.
- Threat-model the feature, not just the happy path.
AI-assisted development is here to stay, and that is a good thing. But "it runs" is not "it's secure." Treat generated code exactly as you would code from an unknown contractor: review it, test it, constrain it, and monitor it in production.
Velizor helps enterprises put these controls into practice — with real-time detection, data protection, and audit-ready governance across the AI stack. Request a demo or explore the AI InfoSec Learning Hub.