Vibe coding, the practice of building software mostly or entirely from AI-generated code with little manual review, is shipping vulnerabilities at industrial scale in 2026. Veracode tested more than 100 large language models on security-sensitive tasks and found 45% of AI-generated code samples introduced an OWASP Top 10 vulnerability, with 86% failing to defend against cross-site scripting and 88% vulnerable to log injection. Georgia Tech’s Vibe Security Radar project traced 35 CVEs in a single month to AI coding tools. And roughly 19.7% of AI-suggested dependencies are non-existent package names that attackers can register and weaponise. The speed is real, and so is the security debt underneath it.
For a publication that tracks how attackers get in, vibe coding matters because it manufactures the entry points. Every unreviewed AI-written endpoint is a candidate for the same injection, auth, and misconfiguration bugs that ransomware crews and access brokers exploit. This piece lays out what the data shows about AI-generated code quality, the supply-chain risk of hallucinated packages, and the real-world leaks that have already resulted, then covers what a security team can actually do about it.
What is vibe coding?
Vibe coding is a development style where a person describes what they want in natural language and accepts the AI’s generated code largely as-is, often without reading it closely or understanding it fully. The term captures a real shift in how software gets built in 2026, from professional developers accepting large AI completions to non-developers assembling entire applications through prompts alone. The productivity gain is genuine. Empirical work across large enterprises found AI-assisted developers commit at three to four times the rate of their peers. The problem is what rides along.
That same research found AI-assisted developers introduce security findings at roughly ten times the rate. Speed and defect rate are climbing together. When the person accepting the code cannot evaluate its security, the AI’s blind spots become the application’s blind spots, and there is no human check between generation and deployment.
How insecure is AI-generated code?
The numbers are consistent across independent studies. Veracode’s test of over 100 models put OWASP Top 10 vulnerabilities in 45% of generated samples, with cross-site scripting defended against in only 14% of relevant cases and log injection in only 12%. OX Security’s analysis found a majority of AI-generated code ships with vulnerabilities. Georgia Tech’s Vibe Security Radar tracked the trend accelerating month over month: six CVEs attributable to AI-generated code in January 2026, fifteen in February, and 35 in March. The models are not getting safer as fast as they are getting adopted.
The recurring weaknesses are the classic ones: injection flaws, missing output encoding, weak or absent authentication, and insecure defaults. These are exactly the bug classes that feed the initial-access economy. An AI-written login endpoint with a subtle auth flaw is indistinguishable, to an attacker, from a human-written one with the same flaw, and there are now far more of them being deployed.
The hallucinated-package supply-chain risk
The most distinctive vibe-coding risk is slopsquatting. Research found approximately 19.7% of AI-suggested dependencies in Python and JavaScript are names that do not exist. When a model confidently recommends a package that was never published, an attacker can register that exact name with malicious code inside, so the next developer who accepts the suggestion pulls the attacker’s package straight into the build. Because AI assistants frequently insert import statements automatically, a developer may not even notice a new dependency was added.
This is a supply-chain attack that the AI itself seeds, a natural extension of the broker-and-dependency risks in the wider software supply chain we cover through initial access brokers and dependency abuse. Unlike typosquatting, which relies on a developer’s typo, slopsquatting relies on the model’s hallucination, which is more predictable and therefore easier to pre-position against.
It is already leaking real data
The Moltbook exposure, found by Wiz researchers on February 1, 2026, is the cautionary case. Its founder built an AI social network entirely with AI coding tools, writing no code himself, and exposed 1.5 million API authentication tokens plus 35,000 email addresses within days of the site’s viral late-January launch. This is what the statistics look like when they reach production: not an abstract OWASP category, but authentication tokens sitting in the open because no one who understood the code reviewed it before it shipped. Exposed tokens and credentials are precisely the raw material that ends up in the stealer-log and credential markets feeding downstream compromise.
What should teams do?
Keep the speed, add the gate. AI-generated code should pass the same static analysis, dependency verification, and security review as any other code, and for vibe-coded projects that review has to be non-negotiable because the author cannot vouch for it. Pin and verify every dependency, and block installs of packages that fail an existence-and-reputation check, which defeats slopsquatting directly. Run SAST and secret-scanning in the pipeline so exposed tokens like Moltbook’s are caught before deployment, not after.
The organisational fix is to treat AI as a fast junior developer whose output is always reviewed, never as a trusted senior whose code ships unread. The ten-times-higher finding rate is manageable if there is a security gate between generation and production. It is a breach waiting to happen if there is not. Static analysis and dependency verification in the pipeline are the containment layer once something does slip through.
Frequently asked questions
What is vibe coding?
Vibe coding is building software mostly from AI-generated code accepted with little manual review, often by people who do not read or fully understand the output. It is fast but pushes the AI’s security blind spots straight into production.
How much AI-generated code is insecure?
Veracode found 45% of AI-generated code samples introduced an OWASP Top 10 vulnerability, with 86% failing to defend against cross-site scripting and 88% vulnerable to log injection. Georgia Tech traced 35 CVEs in a single month to AI coding tools.
What is slopsquatting?
Slopsquatting is registering package names that AI assistants hallucinate. About 19.7% of AI-suggested dependencies do not exist, so an attacker can publish malicious code under those exact names and wait for developers to accept the suggestion.
Has vibe coding caused a real breach?
Yes. The Moltbook AI social network, built entirely with AI coding tools, exposed 1.5 million API tokens and 35,000 email addresses within days of its viral launch in late January 2026.
Does AI coding really produce more vulnerabilities?
Research across large enterprises found AI-assisted developers commit three to four times faster but introduce security findings at roughly ten times the rate, so defect volume rises with speed unless a review gate is in place.
How can teams use AI coding safely?
Treat AI as a junior developer whose output is always reviewed. Run static analysis and secret-scanning in the pipeline, pin and verify every dependency to defeat slopsquatting, and require security review for vibe-coded projects before deployment.
Sources and further reading
- Wiz: Exposed Moltbook database reveals millions of API keys
- OX Security: Why 62% of AI-generated code ships with vulnerabilities
- Cloud Security Alliance: Vibe coding’s security debt, the AI-generated CVE surge
- TechTimes: AI coding agents skip package verification, and attackers are exploiting it
- Kusari: Vibe coding is shipping vulnerabilities, a security team’s guide
- Ransomnews: Initial Access Brokers 2026: ransomware’s supply chain