WhatsApp
AI Empowers Everyone to Build — But Not Everyone to Build Safely
CybersecurityAI DevelopmentVAPTEnterprise Security

AI Empowers Everyone to Build — But Not Everyone to Build Safely

S
Sangeet Chopra
1 September 2026 · 6 min read

AI didn't create a new threat. It removed the one safeguard that used to catch it: a developer who knew what they were doing.

In 2023, Samsung engineers pasted proprietary semiconductor source code into ChatGPT to help debug it. Within weeks, that code — trade secrets the company had spent years protecting — was sitting inside a third-party model's history, outside Samsung's control. Samsung banned generative AI tools company-wide shortly after. The engineers weren't reckless. They were using a tool the way millions of people use it every day. Nobody had told them that pasting internal code into a public LLM is functionally the same as emailing it to a stranger.

That's the pattern worth paying attention to. AI does not need to be malicious to cause damage — it just needs to be used by someone who doesn't understand what it's actually doing under the hood.

The "it compiles, ship it" problem

AI coding assistants are extraordinarily good at producing code that looks correct and runs without errors. That is precisely the danger. A non-technical founder, a marketing team building an internal tool, or a junior developer without a security background can now generate a working login page, a payment form, or a customer database API in an afternoon — without ever learning why certain patterns are unsafe. Independent research from application-security firms studying AI-generated code has repeatedly found the same failure modes showing up at meaningfully higher rates than in human-written code:

  • SQL injection — string-concatenated queries instead of parameterized statements, because the model optimizes for "does this return the right answer in the demo," not "can this be exploited."
  • Hardcoded secrets — API keys, database passwords, and tokens written directly into source files and pushed to public repositories. GitGuardian's annual secrets-sprawl research has tracked a steady rise in exposed credentials inside AI-assisted commits.
  • Broken authentication and access control — endpoints that check whether a user is logged in but never check whether they're logged in as the right user, which is exactly the class of bug behind most real-world API breaches (OWASP calls this Broken Object-Level Authorization, and it consistently ranks among the most common vulnerabilities we find in client audits).
  • No input validation or rate limiting — because the AI was asked to "make the form work," not "make the form survive someone submitting 10,000 malformed requests a second."

None of this is the AI "failing." It's doing exactly what it was asked: produce functional code quickly. Security was never part of the prompt, and it was never part of the person's mental model of what to check for.

Why this is an organizational risk, not just a coding risk

The OWASP Foundation — the same body behind the OWASP Top 10 that security teams have used for two decades — has published a dedicated Top 10 for Large Language Model Applications, covering risks like prompt injection, insecure output handling, and training data leakage. That a globally recognized security standards body felt the need to write an entirely new top-10 list specifically for AI systems tells you this isn't a hypothetical concern — it's already a recognized category of enterprise risk. The organizational exposure shows up in three places:

  1. Shadow AI development. Teams outside engineering — sales ops, HR, marketing — are increasingly building their own internal tools using AI copilots, often without IT or security ever reviewing them. These tools frequently touch real customer or employee data.
  2. Data leakage through the prompt itself. Every piece of proprietary code, customer data, or internal document pasted into a public AI tool is a piece of data that has left your security perimeter — regardless of what the vendor's data-retention policy claims.
  3. False confidence. "The AI wrote it, so it must be fine" is a genuinely dangerous organizational belief. AI-generated code needs the same code review, static analysis, and penetration testing as human-written code — arguably more, since it's being produced faster and reviewed less.

AI empowers everyone. That's exactly why it needs guardrails.

None of this is an argument against using AI for development — the opposite, in fact. AI-assisted development is one of the highest-leverage shifts we've seen in software delivery, and at CyberCure we build AI directly into the platforms we deliver for clients: intelligent automation, AI-vision quality inspection, document processing, and more. The technology is not the risk. The gap between who can now produce code and who understands what makes code safe is the risk.

The organizations getting this right share a common pattern: they let AI accelerate development, but they still route everything through people who understand security, authentication, data handling, and infrastructure — the technical depth that catches what the AI won't flag on its own. AI is a force multiplier for a skilled team. It is a liability for a team that has no one who can tell the difference between code that runs and code that's safe.

What we recommend to clients building with AI-assisted development

  • Treat every AI-generated pull request the same as any other — full code review, SAST/DAST scanning, and dependency checks before merge, no exceptions for "the AI wrote it."
  • Never paste proprietary code, credentials, or customer data into public AI tools without an enterprise agreement that guarantees data isn't retained for training.
  • Maintain an inventory of every internal tool built with AI assistance, including the ones built outside the engineering team — you cannot secure what you don't know exists.
  • Run a VAPT (vulnerability assessment and penetration test) on any AI-assisted application before it touches production data, the same as you would for any other release.
  • Train non-developer teams building with AI copilots on the basics: what a secret is, why it can't live in source code, and when to call in someone with security expertise.

AI is not the threat. An organization that lets anyone ship code without anyone who understands what "safe" looks like — that's the threat. The fix isn't slowing down AI adoption. It's making sure the humans directing it still have the depth to know what they're looking at.

Share this post
Back to Blog
Why AI-Generated Code Is a Security Risk Without Technical Oversight | CyberCure