Vibe coding is safe for prototypes and throwaway tools, and risky for anything that holds real user data. The code often works but has not been reviewed, and studies find security flaws in a large share of AI output. Before you ship, review authentication and data access, scan for secrets and vulnerable dependencies, and keep the assistant away from production.
Key takeaways
- The risk comes from skipping review, not from using AI.
- Veracode found security flaws in 45 percent of AI code tasks it tested.
- Common failures are exposed secrets, missing access control and packages that do not exist.
- A short checklist and three scans cover most of the risk.
What is vibe coding?
Vibe coding is building software by describing it in plain language to an AI model and accepting the code with little review. Andrej Karpathy coined the term in February 2025, describing it as giving in to the vibes and forgetting the code exists.
Is vibe coding safe for production?
Not by default. Reports show what happens when it ships unchecked. In May 2025, reports said 170 of 1,645 web apps built with the Lovable platform had a flaw that let anyone access personal information. In July 2025, a founder reported that Replit's AI agent deleted a database despite being told not to change anything.
Veracode's 2025 report tested 80 tasks across more than 100 models and found vulnerabilities in 45 percent of cases. Its follow up finding was that security did not improve over time, even as the code got more correct.
These are one platform's reports and one vendor's study, so read them as signals, not guarantees.
What goes wrong most often?
Four failures show up again and again.
- Missing access control. The app works for you, but any user can read other users' data.
- Secrets in code. API keys end up in the repository or the front end.
- Invented packages. A paper testing 16 models found hallucinated package names averaging at least 5.2 percent for commercial models and 21.7 percent for open source ones. Attackers can register those names.
- Over-trusted agents. An agent with write access to a database can destroy it.
How do you make a vibe coded app safe enough to ship?
Review the parts that matter, then scan the rest.
- Read every route that touches users, payments or data. Check that each one verifies who is asking.
- Search for secrets and move them to environment variables.
- Check each dependency exists, is the real package, and has no known CVEs. You can paste a
package.jsonorrequirements.txtinto the free scan at gold.safeguard.sh/scan. It runs in your browser, needs no login and covers up to 100 dependencies. - Give the agent a development database, never production.
- Add a static analysis scan to your pull requests. See how to scan AI generated code.
Checklist before you launch
- Every data route checks the caller
- No secrets in code or history
- Dependencies verified and scanned
- Agent has no production access
- Backups exist and have been restored once
- Someone other than the assistant has read the auth code
Frequently asked questions
Can beginners vibe code safely?
For personal projects, yes. For apps with other people's data, get a review from someone who knows security first.
Is AI code less secure than human code?
One analysis summarized on Wikipedia found AI co-authored pull requests had about 1.7 times more major issues. Treat it as a reason to review, not proof for every project.
Do bigger models write safer code?
Veracode reported that larger models did not perform significantly better than smaller ones.
What is the one check to do if I only have time for one?
Check that every route returning user data verifies the caller. Missing access control does the most damage.
Sources
- Vibe coding overview and incident summaries: https://en.wikipedia.org/wiki/Vibe_coding
- Veracode 2025 GenAI Code Security Report press release: https://www.veracode.com/press-release/ai-generated-code-poses-major-security-risks-in-nearly-half-of-all-development-tasks-veracode-research-reveals/
- Package hallucination study: https://arxiv.org/abs/2406.10279
- Safeguard Gold dependency scan: https://gold.safeguard.sh/scan
Last reviewed: 19 October 2026.
