Vibe coding may be unstoppable—but here’s how to rein in the risks

Leo thought he’d found an amazing shortcut to software development.

Earlier this year, Leonel Acevedo, founder of a lead-generation marketing software startup and a rookie coder, launched an SaaS (software-as-a service) application written entirely with AI. He never directly touched the code himself. Instead, he relied on plugging prompts, or instructions, into an AI tool and let the system do the work. Then, he proudly told the world about his accomplishment in a post on X.

Big mistake.

The risks from third parties and software supply chains are rising. Know your vulnerabilities—in seconds.

Two days later, his creation was under assault. Hackers maxed out his application programming interface (API) keys, essentially overrunning the system that controls how apps talk to each other. They bypassed his subscription systems, which are meant to block anyone who isn’t a paying user. Then they started creating random entries in his database, showing they could tinker with the data however they wanted.

“Half the internet is trying to break my app just for sharing publicly a mistake I made,” wrote Acevedo, who quickly shut down the app. “You know what's the worst thing? They are doing it for fun.”

The implosion was quick and instructive.

Vibe coding, the practice of using AI models to create full applications from natural language prompts, has become a mainstream way to build software. “The bar to coding is now lower than it ever has been,” said Google Brain founder Andrew Ng in a talk at Snowflake’s Build conference on Monday. But the rapid development times and low costs that make vibe coding so seductive also create risks by pushing unreviewed, automatically generated code into pipelines with little or no IT security review. As a result, more individuals and organizations are destined to find themselves using software containing record numbers of glitches that could bog down operations and vulnerabilities that hackers can easily discover and exploit.

“If you’re a good coder, you can get a lot of help from AI because you know what you’re looking for and how to nudge it,” said Erlend Oftedal, senior technical director at RSA Conference (RSAC) Research covering AppSec. “But if you’re a bad coder, it’s only going to make things worse because you don’t understand how to review what you’re developing. You’re just churning stuff out.”

Vibe coding: the good vibes, the bad, and all that venture capital

While the term vibe coding is new, recently coined by AI researcher Andrej Karpathy and just last week named the word of the year, the practice of using AI to guide code creation is not.

“The bar to coding is now lower than it ever has been.”
Andrew Ng, founder, Google Brain

In fact, it dates back to the early 2010s, when researchers started using machine learning’s neural networks—software modeled loosely on how the brain learns—for code completion and program synthesis. That paved the way for tools like DeepCoder in 2016 and Google’s early sequence-to-sequence models, and by the early 2020s, systems such as OpenAI’s Codex and GitHub Copilot were bringing AI-assisted coding into the mainstream.

Now comes the latest in that long push, with generative tools like ChatGPT, GitHub Copilot, Anthropic’s Claude, and Cursor (what Acevedo used) making coding even faster and easier—and more appealing to business enterprises. Last year, more than 90% of developers at medium and large companies acknowledged using AI to generate code, according to a Checkmarx/Censuswide survey. And 30% to 40% of organizations now actively promote vibe coding, according to a 2024 GitHub survey, presumably to cut costs and accelerate software development lifecycles (SDLs).

To be fair, software companies have always shipped code before it’s ready in the interest of beating competitors to market. New operating systems and applications notoriously contain numerous operational and security bugs that can take enterprises months to patch and correct. The risky practice of slow patching has become endemic. But what worries security experts now is that AI-assisted coding may produce vulnerabilities at such scale that many will never be identified or fixed.

In fact, it’s already happening. In a recent Veracode study, 45% of more than 100 AI-generated code samples tested failed security tests and introduced critical web application security risks.

Such concerns aren’t stopping vendors and investors from jumping on the vibe coding bandwagon. In fact, vibe coding tool startups have recently attracted significant venture capital funding, including Cursor-maker Anysphere, which raised $900 million; Replit, which secured $250 million; and Lovable, which raised $200 million.

While not purely security-focused, many of these tools include features like security scanning, which checks code for exploitable flaws, and automatic API key protection, which detects and blocks exposed credentials before they can be misused. However, they tend to stop at detecting issues rather than validating whether the underlying protections actually work.

Vibe coding’s biggest risk-takers

The surge of new tools is also shaping who’s taking the biggest risks with vibe coding. Pieter Danhieux, co-founder and CEO of Secure Code Warrior, said startups tend to embrace it more aggressively, while enterprises stick to assisted coding with a human still reviewing the output.

“Coding an application containing sensitive data without the skills to secure it is a recipe for disaster.”
Pieter Danhieux, co-founder and CEO, Secure Code Warrior

“I think vibe coding is widely adopted in smaller startups where there’s a brilliant technical founder who wants to move fast, and they basically start building applications and deploying them without actually looking at them,” Danhieux said.

Larger enterprises, he explained, are more cautious. Their developers might use AI to create small, reusable pieces of code, but they still check the results before deployment.

While seemingly safer, Danhieux warned those checks may provide nothing more than a false sense of safety.

Secure Code Warrior research found 86% of developers are not prioritizing secure coding, and several popular LLM tools struggle to solve secure coding challenges, especially those in more subjective categories like authentication and security misconfiguration, Danhieux said.

Vibe coding’s blind spots and sensitive data

Language and vulnerability types make a difference, Danhieux noted. Models tend to catch simple mistakes such as spelling errors in code or obvious problems that follow a clear pattern. But they struggle with deeper issues that involve how a system is designed, including whether an app properly checks who is allowed to do what. That creates a blind spot because AI-generated code can look clean on the surface but still contain weaknesses that put users at risk.

“That’s the danger. You get code that looks convincing, runs in production, and has never been through a proper review. The bugs hide in plain sight.”
Erlend Oftedal, senior technical director, RSA Conference (RSAC) Research

That same dynamic is beginning to show up outside application development as well, in abuse patterns like VibeScamming, where natural‑language AI tools are deliberately used to generate and rapidly iterate phishing pages and scam infrastructure that look legitimate but bypass traditional trust signals.

Savvy developers might catch such hidden vulnerabilities, but novices probably wouldn’t have the requisite experience or intuition to do so, Danhieux said. He points to a recent case underscoring the risk. Last spring, a comprehensive scan of 1,645 projects created on the Lovable AI coding platform turned up 303 vulnerable endpoints across 170 applications built with the tool, or about 10% of analyzed projects.

“Coding an application containing sensitive data without the skills to secure it is a recipe for disaster,” he said.

Currently available technology only complicates matters, as it isn’t yet up to the task of catching security problems, especially when code is being generated in high volumes. AI-powered tools may soon help with detection, but that still won’t be enough, said RSAC’s Oftedal.

“Most companies think their open-source governance covers AI—it doesn’t,” he said. “That’s the danger. You get code that looks convincing, runs in production, and has never been through a proper review. The bugs hide in plain sight.”

Oftedal noted that developers themselves can help by prompting AI to write code more securely or to mimic the security approaches taken by other respected developers. They can also instruct AI to run certain tools to check their work before moving it along, he said.

Vibe check: the steps CEOs and security leaders need to take now

For experts like Danhieux and Oftedal, the path forward starts with ensuring visibility and oversight.

“A functional ‘secure by design’ initiative… is the antidote to out-of-control technical debt and costly late-stage security measures.”
Danhieux

Just like enterprise leaders need to know the vast array of computers, tablets, and other endpoints on their computer network, with real-time data on operating systems and patch management needs, chief information security officers need to know which AI coding models their developers are using and set boundaries, they said. Cheaper, older models could produce less-secure code, and without oversight—a software bill of materials (SBOM), for starters, to list an app’s components and third-party dependencies—organizations often have no idea what’s running in their environments.

“Maybe we’re going to go to a situation where there will be guardrails for developers saying, you can use Claude version three and above, and you can use ChatGPT-4 and above, but nothing below, because we know the code coming out of that will be insecure,” Danhieux said.

Training developers is the other half of the equation. They need to learn prompt engineering and how to guide models toward safer outputs, just as writers learn to coax better prose from AI. Until enterprise-wide tools can enforce policy inside every integrated development environment, awareness and training are the stopgaps.

But for Danhieux, the guiding principle is secure by design, which means building security into software from inception.

"A functional ‘secure by design’ initiative, upheld by security-proficient developers, is the antidote to out-of-control technical debt and costly late-stage security measures,” he said. “Threat actors already have an advantage over enterprise security teams, but prioritizing security at the earliest stage in the SDL can help swing the pendulum back in our favor."

The longer-term concern, Danhieux warned, is that vibe coding will continue pushing software creation out of enterprise development pipelines and into living rooms, leading to more vulnerabilities that could expand the threat landscape. Professional developers may not be perfect, but they still move through required checks. Novice coders do not, and as their software spreads, the attack surface grows and trust in the system shrinks.

Danhieux argues that fixing this gap is critical before the problem scales even further.“If we don’t solve this security problem, I think we’re going to be worse off in a few years,” he said.

Oftedal echoed the urgency.

“The bugs are already out there,” he said. “If you are not addressing this now, you are already behind.”