The AI Code Slop Problem Is Getting Worse
AI code slop was a concern in 2025. In 2026, it is a crisis. The term describes unreviewed code generated by AI coding assistants that looks clean, compiles without errors and appears to function correctly, but contains security vulnerabilities that the builder never sees and the AI never mentions.
The volume of AI-generated code shipping to production has increased dramatically. GitHub reported that over 40% of all code committed to the platform in 2025 was AI-generated. That number has only grown. Tools like Cursor, Replit Agent, Claude Code, Bolt and Lovable have lowered the barrier to shipping software to zero. Anyone with an idea and a credit card can build and deploy a production application in an afternoon.
The security implications are significant. When millions of applications are built by people who have never written a line of code before, using AI tools that prioritize functionality over security, the result is an enormous expansion of the internet's attack surface.
What Our 2026 Data Shows
We have audited hundreds of AI-generated applications in the first quarter of 2026 alone. The data is consistent with what we published in our State of AI Code Security report, but the trend lines are moving in the wrong direction.
| Vulnerability | Prevalence | Typical Impact |
|---|---|---|
| Hardcoded API keys/credentials | 72% of apps | Full account takeover |
| SQL injection | 58% of apps | Database compromise |
| Missing auth on admin endpoints | 64% of apps | Unauthorized access |
| Insecure direct object references | 53% of apps | Data leakage |
| Overly permissive CORS | 71% of apps | Cross-origin attacks |
| Missing rate limiting | 84% of apps | Brute-force attacks |
These numbers have not improved since 2025. In some categories, they have gotten worse. The reason is simple: the people building with AI tools are building more complex applications, but the AI tools have not meaningfully improved their security awareness.
Why AI Tools Generate Insecure Code
AI coding assistants are trained on the internet's code. That includes millions of tutorials, Stack Overflow answers, blog posts and open-source repositories. Much of that training data contains insecure patterns. The AI learns to generate code that works, not code that is secure, because the training data overwhelmingly rewards functionality over security.
- Security is a non-functional requirement
- When you tell an AI to build a login page, it builds a login page. It does not think about brute-force protection, session fixation, credential stuffing or password storage best practices unless you specifically ask. Most vibe coders do not know to ask.
- AI optimizes for the happy path
- AI-generated code handles the expected use case well. It rarely handles the adversarial use case at all. What happens when a user submits a 10MB file to your upload endpoint? What happens when someone sends a SQL string as a username? The AI does not ask these questions.
- Context windows create blind spots
- AI assistants see the code in their current context window. They do not see the full architecture. They do not know that the API endpoint they just created is accessible without authentication because the middleware configuration is in a different file they were not shown.
- Hallucinated dependencies
- AI tools sometimes reference packages that do not exist. Attackers have started registering these hallucinated package names and publishing malicious code under them. If your AI suggests installing
express-auth-validatorand that package was published by an attacker two weeks ago, you now have a supply chain compromise.
The Scale Problem
The real danger of AI code slop is not any individual vulnerability. It is the scale. When one developer writes insecure code, one application is at risk. When millions of people use AI tools to generate the same insecure patterns across millions of applications, the entire internet's security posture degrades.
We are seeing the same vulnerabilities repeated across hundreds of applications because the AI tools generate the same patterns. The same insecure database query. The same missing authentication check. The same hardcoded credential. It is as if every AI coding tool is teaching the same bad habits to millions of builders simultaneously.
What You Can Do About It
If you are building with AI coding tools, you need to acknowledge a fundamental truth: the AI does not make your code secure. It makes your code functional. Security is your responsibility.
Get an audit before you launch. A professional code audit catches the vulnerabilities that AI tools introduce. Our Quick Audit starts at $1,500 CAD and covers the most common AI code slop patterns.
Use security-focused prompts. When asking AI to generate code, explicitly request input validation, parameterized queries, authentication checks, rate limiting and proper error handling. The AI will not add these by default.
Never hardcode secrets. If your AI assistant puts an API key directly in your source code, move it to an environment variable immediately. Better yet, use a secrets management service.
Verify every dependency. Before installing any package your AI suggests, verify it exists on the official package registry and check its download count, maintenance status and publish history. Our guide on verifying AI-suggested packages walks through this process.
Test authentication on every endpoint. Open your browser's developer tools, copy an API request, remove the authentication token and send it again. If the endpoint responds with data instead of a 401 error, you have a critical vulnerability.
The Industry Must Respond
AI code slop is not going away. The tools are too useful. The productivity gains are too significant. People are going to keep building with AI assistants and the volume of AI-generated code shipping to production will continue to grow.
What needs to change is the assumption that AI-generated code is safe by default. It is not. Every application built with AI tools needs a security review before it handles real user data. The cost of that review is a fraction of the cost of a data breach.
Read our full analysis in What Is AI Slop? for a deeper look at the problem, or visit our AI code audit service page to get your application reviewed.