Vibe-Coded App Risk Report — Week of Oct 26, 2026: Is Your AI-Built App Safe?

“Is my AI-built app safe?” is a fair question for any business owner to ask, and this week gave several practical answers. Most of the news wasn’t about the apps themselves. It was about the tools and laptops used to build them, the defaults underneath them, and who looks after them once they’re running.

Here are the stories from the week of October 26 that matter to a business running on an app built with Lovable, Bolt, Replit, Cursor, Claude Code or a freelancer using them, with one action for each.

A tampered Claude Code extension reached up to 5,000 Cursor users

On October 29, Cursor published a security notice about a malicious version (2.1.313) of the Claude Code extension that appeared in Cursor’s own extension marketplace. Cursor says the version “was not developed by Anthropic.” It was available for about eleven hours, from roughly 2:27 a.m. to 1:35 p.m. Pacific time that day, and reached at most 5,000 users. Cursor says it will contact them directly.

The tampered extension tried to add an attacker’s SSH key (labeled svc-telemetry) to the computer, set itself up to keep running, and send back details about the machine, including the names of environment variables that looked like credentials. Cursor notes that one step usually needed admin rights and likely failed on most machines, but the key, the persistence and the report back could still happen. Anthropic has since published a clean version, 2.1.314.

What it means for a business: your app is only as safe as the laptops that build it. If an employee or freelancer uses Cursor, a compromised editor on their machine can reach the same keys and accounts your app uses.

Action: ask anyone who works on your app in Cursor whether they installed or updated the Claude Code extension on October 29. If they did, they should follow Cursor’s notice in full. Cursor warns that uninstalling the extension alone may not be enough. Then replace any keys that machine could reach.

AI coding agents took orders from a file in the Downloads folder

Forever Security found a malicious instruction file (AGENTS.md) in a customer’s Downloads folder, written to look like ordinary engineering notes. Coding agents follow the instruction files in whatever folder they start in. In the researchers’ tests, Codex followed the planted file 90% of the time, and a similar CLAUDE.md file got Claude Code to send its traffic, and its login token, to a server the researchers controlled. The researchers say there is “nothing here for OpenAI or Anthropic to patch.”

What it means for a business: this is a habit problem, not a software bug, so a written rule fixes most of it.

Action: agree one rule with whoever builds your app: start coding agents only inside the project’s own folder, never in Downloads or the home folder, and keep approval prompts switched on.

Supabase stops exposing new tables automatically, starting today

From today, October 30, new tables on existing Supabase projects aren’t reachable from your app until someone grants access. Existing tables keep working, and row-level security behaves as before. The risk is the quick fix someone reaches for when the next AI-built feature shows “permission denied.” We covered the details, and the safe pattern, in Supabase “permission denied” after Oct 30.

Action: tell your AI tool, in its project instructions, that every new table needs a grant, row-level security and an access policy in the same change.

26,400 working keys found in public code

A scan of 7,256,531 public repositories on Gitee, China’s largest code-hosting site, found 26,400 unique credentials that still worked, published on MongoDB’s blog on October 28. The largest group was Alibaba Cloud keys, and the long tail included Stripe, Slack and AI service keys. Nearly half dated from 2025.

What it means for a business: a key pasted into code keeps working until someone replaces it. To get something working fast, AI tools and freelancers can end up putting keys straight into the code.

Action: ask where your app’s keys live. They belong in your hosting platform’s secret settings, not in the code, and someone should own replacing them when a developer leaves.

Is my AI-built app safe if an employee built it?

Two reports this week put numbers on that question.

Pave, Quickbase’s AI app builder, surveyed 527 business leaders and 5,177 employees in the US in September. It’s a vendor’s survey, so read it with that in mind, but the numbers are striking. 24% of employees said they had built an app with AI that they use for work. 53% of organizations had found at least one app that IT and leadership didn’t know existed. 57% of leaders require a security assessment, but only 32% of builders have someone who can help them do one. Only 14% of builders said someone had been named to maintain their apps.

Veracode’s fall GenAI Code Security Report, released October 29, tested 18 AI models. The average share of tasks producing secure code rose from 56% to 62%, the biggest jump since the series began. That still means roughly 4 in 10 tasks produced vulnerable code. The best models scored 73%.

What it means for a business: AI-written code is getting better, but not good enough to skip a review, and many employee-built apps have no owner at all.

Action: make a list of the apps your staff have built, who uses each one, what data it touches, and who fixes it when it breaks.

This week’s checklist

  1. Ask anyone using Cursor whether they installed Claude Code extension 2.1.313 on October 29. If so, follow Cursor’s notice and replace keys.
  2. Set the rule: coding agents start only in the project folder, with approval prompts on.
  3. Add “grant + row-level security + policy for every new table” to your AI tool’s instructions for Supabase apps.
  4. Confirm your app’s keys live in secret settings, not in the code.
  5. List every app your staff have built with AI and name an owner for each.

If you’d like a second pair of eyes, our free Forward Scan shows the 3 biggest risks in an AI-built app.

Sources

Does your business run on an AI-built app?

Our free Forward Scan is a 20-minute video walkthrough of your app’s 3 biggest risks, delivered in 2 business days.

← All posts

Free
2 days
Get a free Forward Scan