90% of AI-Built Web Apps Have Vulnerabilities | The Dangers of Vibe Coding Revealed
機械翻訳 / Machine-translated
A
AI Friends
5日前・13分で読める
機械翻訳 / Machine-translated
@aifriends
AI Friends（https://aifriends.jp）のクロスポスト公式アカウント。AIツールの紹介・使い方・できることを、中学生でもわかるやさしい日本語で届けます。
What you'll learn from this article
On October 7, 2026, a research team including Microsoft and the Singapore AI Safety Institute announced significant findings. After examining web applications automatically generated by AI, they discovered vulnerabilities in 91% of the apps.
The study looked at 200 publicly available web apps built through vibe coding. Vibe coding refers to the approach of creating apps simply by giving instructions to AI. A total of 1,186 vulnerabilities were found, of which 65.77% were classified as critical or high severity.
According to reporting by ITmedia, the research team randomly sampled 200 apps from 984 apps built with AI tools such as Claude and Lovable for the study.
Vibe coding is an approach where anyone — even without programming expertise — can simply tell an AI "I want an app like this" and have a finished app produced. Rather than writing code yourself, you convey a general idea to the AI and let it do the building.
For example, just saying "an SNS app where users can log in and share posts" prompts the AI to automatically generate everything from database design to front-end screens. Development time has been cut from the traditional several weeks down to just a few hours.
As convenient as this is, it comes with a problem: AI produces working code, but it doesn't fully account for security.
The three most common vulnerabilities discovered in this study were as follows.
The first was broken access control, accounting for 339 cases — 28.58% of the total. Data that should be inaccessible becomes accessible to anyone. For instance, users could edit someone else's posts without permission, or ordinary users could access administrator-only pages.
The second was vulnerability to injection attacks, with 199 cases at 16.78%. By entering malicious code into an input field, an attacker can take over the system. SQL injection is a prime example, where an entire database can be exfiltrated.
The third was authentication failures, making up 176 cases at 14.84%. Weak login functionality allows entry without a password or makes it easy to impersonate another user.
A separate study reported by SecurityWeek found that among 434 vulnerabilities, 23 were rated critical — and 11 of those involved passwords or API keys hardcoded directly into the code.
The research team identified three categories of root causes.
The most common is knowledge defects, accounting for 63.4%. These arise because the AI is unaware of security best practices. For example, rather than following the basic rule of encrypting passwords before storing them, the AI saves them in plaintext — meaning in a directly readable form.
The second is objective defects, at 23.1%. When user instructions are vague or say nothing about security, the AI simply produces code that works, leaving security as an afterthought.
The third is memory defects, at 13.5%. During a conversation, AI can forget earlier instructions. For instance, even if you specify at the outset that a page should only be accessible to administrators, the AI may later forget that restriction and generate code that allows anyone to access it.
This issue isn't confined to Japan. According to Veracode's 2026 report, 44% of AI-generated code contains dangerous vulnerabilities, and the average security test pass rate is just 56%.
Particularly concerning are the extremely low success rates for cross-site scripting at 15% and log injection at 12%. By contrast, SQL injection scores 83% and encryption algorithms 87%, revealing a clear divide in what AI handles well versus poorly.
In the second half of 2025, vulnerability reports totaled 18 over seven months, but that number surged to 56 in just the first three months of 2026 — with 35 reported in March 2026 alone, surpassing all of 2025. As vibe coding spreads, the problems are growing rapidly too.
The study also evaluated the effectiveness of several countermeasures.
The first is improving prompts. Simply adding "please build a secure app suitable for production use" to AI instructions reduced the vulnerability recurrence rate from 25.7% to 11.0%. Since AI won't consider what it isn't told to consider, clearly specifying security requirements is essential.
The second is adding security skills. Incorporating additional security-enhancement capabilities into the AI tool brought the vulnerability rate down to 11.9%, showing that improvements on the AI tool side are also effective.
The third is human review. An article on @IT notes that experts insist AI-generated code must always receive a final check by a human. Yet currently, only 12% of organizations apply the same security standards to AI-generated code as they do to traditionally written code.
The use of vibe coding is growing in Japan as well. GitHub Copilot surpassed 20 million users in mid-2025, and Google reported that approximately half of all new code written in early 2026 was AI-generated.
While development speed is increasing, security practices are failing to keep pace. Small and medium-sized businesses and startups in particular are increasingly publishing apps built with AI tools without having a dedicated security professional on hand.
The findings of this study highlight the risks hiding behind convenient tools. AI is a powerful assistant, but the final judgment on security must rest with humans. Anyone using vibe coding should ensure their work is reviewed by a security professional.
This article is a cross-post from AI Friends.
