Your coding agents are still shipping your secrets: the state of the VS Code Marketplace
I checked all 154,420 extensions in the VS Code Marketplace. Almost 1 in 100 ships Claude Code's local settings file, and about 30 AI tools leave their files behind.

In May I wrote about .claude/settings.local.json (blog here), the file where Claude Code remembers every command you clicked "allow always" on. Lakera found it in 0.92% of new npm packages, and about 1 in 13 of those files held real credentials. My own scan found it on about 1 in every 1,000 websites.
I wanted to know what happens in the place where developers actually work: their editor. VS Code extensions run on developer machines full of API keys, SSH keys and cloud logins. And the people who publish them hold a token that can push an update to every user.
So I checked the whole VS Code Marketplace.
I'll go into the details below, but here are the highlights:
- Almost 1 in 100 extensions ships Claude Code's local settings file. About 4,500 extensions ship files from some AI coding tool, and some of them have millions of installs.
- Live keys still ship. I found them in settings files, MCP configs, a Claude skill and plenty of
.envfiles. - 17% of the shipped allowlists let the agent publish a new version of the extension without asking anyone.
- The settings file peaked in early 2026 and dropped fast after Claude Code's auto mode became available to everyone. The old files are still out there, though.
What this means for security teams and developers is that agent files are now part of the software supply chain, and the default publishing tools don't check them.
Why this matters
Extensions run with the developer's full permissions, with no sandbox, and VS Code updates them automatically. Whoever holds the publisher's token can run code on every machine that installed the extension. Agent files make that token easier to get: sometimes the file contains it, and even a harmless one shows how the extension is built and published. And if an agent's local files made it into a release, nobody checked what went into that package.
What I did
- I listed every extension in the Marketplace through its public API, which came to 154,422 extensions. Two of them failed to download, so the scan covers 154,420.
- An extension package is a zip file, and a zip keeps its table of contents at the end. I downloaded only that part, so I saw the name, size and date of every file in every package without downloading the packages themselves.
- I looked for files from about 30 AI coding tools: Claude Code, Copilot, Cursor, Codex, Gemini, Cline, Kiro, Aider and others, plus MCP configs (the files that connect agents to external tools). I counted only files that belong to the extension itself, not to its bundled dependencies.
- I downloaded only those files, plus any
.envfiles as a comparison, and scanned them with TruffleHog with verification on. TruffleHog is an open-source secret scanner that can check whether a key still works. - I read every settings file and MCP config that matched a credential pattern by hand.
The whole scan took about 45 minutes. I didn't scan the code of the extensions, only the agent files and .env. So the secret numbers below are the minimum, not the full picture.
The numbers
1,374 extensions (0.89%) ship .claude/settings.local.json. That's almost exactly the rate Lakera saw on npm (0.92%), and about nine times the rate I saw on websites.
The settings file is only one example. 4,504 extensions ship some agent file. CLAUDE.md is in 1,181 of them, AGENTS.md in 989, Copilot instructions in 604 and Cursor files in 426. 605 extensions carry files from two or more agents, so the developer tried several tools on the same project.
It's not only hobby projects either. 46 of these extensions have more than 10,000 installs, and 12 have more than 100,000. The biggest one is an AI coding assistant with 2.9 million installs. Microsoft's CodeTour and one of the most popular icon themes are on the list too. Their files are harmless allowlists, but nobody meant to ship them.
And it's not only AI tools: 83% of the affected extensions are ordinary themes, linters and language support.
It peaked, and then auto mode happened
The timeline was the part I didn't expect.
No extension last updated before 2025 ships the settings file. Then it grows fast. By early 2026, about 4% of extensions updated each month ship it, and it stays around 3 to 4% until June 2026. Then it drops:
- In July 2026 it's 2.0%.
- In September it's 1.0%.
- In October it's 0.5%.
CLAUDE.md didn't drop the same way, so people didn't stop using Claude Code. They stopped clicking "allow always".
The drop starts the month Claude Code's auto mode became available to everyone. In auto mode a classifier decides which commands are safe, so the settings file no longer collects approvals. It's a correlation, but the timing is tight.
Side note for red teamers: with auto mode, nobody clicks "allow" anymore. In most setups the classifier is now the only thing between a prompt injection and the shell. If the injected command looks safe enough to the classifier, it runs. So if you assess AI coding setups, test the classifier too, not only the agent.
That's good news for new extensions. But the old files are still there. Every extension that shipped the file and was never updated keeps shipping it to every new install.
Secrets
Credentials in settings files are rarer here than on npm. 9 of the 1,374 extensions have real credentials in the file, and TruffleHog confirmed 3 keys as live. On npm it was about 1 in 13 files. Here it's closer to 1 in 150.
The few cases that exist are interesting:
- In one file, the allowlist shows the developer asking the agent to split two API keys into chunks and base64-encode them. That's how you hide a key from a scanner. Both keys still work.
- In another, the settings file points Claude Code at a third-party model provider, with the API key in the
envblock. - One extension has what looks like a Marketplace publish token inside a
vsce unpublishcommand (vsceis the tool that packages and publishes extensions). TruffleHog didn't flag it. I didn't test it, but if it's live, it can push an update to every user of that extension.
MCP configs leak at a higher rate. 7 of the 160 extensions that ship an MCP config have a key written directly in it: Anthropic, OpenAI, ClickUp, Context7, GitBook and two smaller services. None of them verified as live, and 4 of the 7 key types have no TruffleHog detector at all. Most configs do it right and use ${env:...} references instead.
Claude skills are a new place to leak. One extension ships a skill whose reference docs contain a live GitHub token.
Scanners don't know these formats yet. TruffleHog missed the probable publish token, an Atlassian token, a MiniMax key, an MCP vendor's own API keys and an LLM proxy key. I found them by grepping and reading.
And .env is still the biggest source, just like when Wiz looked at extensions in 2025. 1,093 extensions ship a real .env file (not an example), and most of the 77 live secrets I found came from those. In extensions updated after Microsoft added secret scanning to the Marketplace in October 2025, 9.3% of shipped .env files still contain a working key. Some of the older keys were probably rotated, so this doesn't prove the scanner fails. It does show that live keys still get published.
What developers let their agents do
The settings files have one more use. Together they hold 12,134 permission rules that developers approved for their agents, one click at a time. Nobody designs these lists. They grow every time someone gets tired of the prompt. So they show what developers really allow, not what a policy says.
- Only 3 files have a deny rule. 62% of all rules contain a wildcard.
- Only 8 extensions allow every shell command outright with
Bash(*). But 252 (18%) allow something likeBash(node:*)orBash(python3:*). It looks narrow, butnode -eruns anything, so in practice it's the same as allowing everything. - 231 (17%) let the agent publish without asking:
vsce publish,ovsx publish,npm publishorgh release create. 144 allowgit push. - 81 extensions allow both. In those projects, an agent that reads a prompt injection from an issue, a README or a web page could build, push and release an update to every user, and nobody would see a single prompt.
- Smaller groups allow
sudo,rm -rf, reading.envfiles, SSH to servers and even reading the macOS keychain.
The permission prompt trained people to say yes. With auto mode, a classifier now makes most of these decisions. These files are a snapshot of what humans approved before it did.
Other things in the packages
Settings files aren't the only thing agents leave behind:
- 13 extensions ship Claude Code worktrees, which are full copies of other branches of the repo: 1,447 files in total, up to 12 MB each.
- 21 extensions ship SpecStory chat histories, 281 full agent conversations in total. SpecStory is a tool that saves your AI chats into the project. One extension ships almost 5 MB of them.
- 17 extensions ship Aider chat histories (Aider is an open-source coding assistant).
- Others ship agent plans and roadmaps, Serena and Claude memory files, and Cursor debug logs.
These had very few secrets. But they hold unreleased code, internal discussions and local usernames, which is plenty for anyone doing recon.
Why it happens
vsce packages everything in the extension folder unless .vscodeignore excludes it. It doesn't read .gitignore. I found one allowlist with the commit message "add .claude/settings.local.json to gitignore", and the file shipped anyway.
vsce already has a default ignore list. It skips .git, .github, .DS_Store and lockfiles, which is why only 8 extensions ship a .git folder. It also refuses to package a .env file unless you pass a flag. No agent folder is on that list.
What to do
If you publish extensions:
- Add
.claude/,.cursor/,.codex/,.specstory/,.aider*and any MCP config to.vscodeignore. - Run
vsce lsbefore publishing to see exactly what goes into the package. - If a key ever shipped, rotate it. Publishing a new version doesn't remove the old one.
- Look at your allowlist. If it has
node:*next tovsce publish, the agent can release on its own.
For Microsoft: add agent folders to vsce's default ignore list. It's a one-line change, and it would fix most of this for everyone. The Marketplace scanner should also look at agent files and MCP configs.
For agent vendors: keep local state out of the project folder, or add the ignore entries the first time the tool runs. Auto mode shows how well a product change works: after one release, new settings files dropped from about 3 to 4% of extensions to 0.5% in four months.
Where VibeGuard fits
I work at Legit Security, and this is the problem our VibeGuard product is built for. It runs on the developer's workstation and hooks into Claude Code, Cursor, Copilot and Codex. Three parts of it map directly to what I found:
- The policy doesn't grow one click at a time. The security team writes the rules, and VibeGuard checks every shell command, MCP call and built-in tool call against them before it runs.
vsce publishorgit pushcan require a human approval, or a fingerprint, in every project. - The agent never sees the key. It gets a placeholder like
VG_SECRET_<id>, and VibeGuard swaps in the real value just before the command runs. So what ends up in an allowlist, a chat history or a skill file is the placeholder. The built-in rules also catch commands that base64-encode key material, like the one in the first example above. - You can see what's installed. VibeGuard lists the agents, MCP servers, skills and plugins on every workstation, so MCP configs and skills stop being a blind spot.
It doesn't check what goes into your .vsix, so you still need the .vscodeignore entries. But a leaked allowlist with placeholders in it is a much smaller problem.
We spent years teaching people to put .env in .gitignore. Now every AI tool brings its own dot folder, vsce ships all of them, and we start teaching all over again :)
Want to learn more? Register here for Kirill's webinar on Thursday, December 3rd.
Legit Security needs the contact information you provide to us to contact you about our products and services. You may unsubscribe from these communications at any time. For information on how to unsubscribe, as well as our privacy practices and commitment to protecting your privacy, check out our Privacy Policy.
More insights for modern AppSec teams

Your coding agents are still shipping your secrets: the state of the VS Code Marketplace
I checked all 154,420 extensions in the VS Code Marketplace. Almost 1 in 100 ships Claude Code's local settings file, and about 30 AI tools leave their files behind.

Why I Joined Legit
We believe that security tools had become very good at identifying problems, but that developers were still left with the hardest part: understanding those findings and turning them into secure code changes.



See it in your stack
Every prompt, every commit, every deploy, covered the moment you're ready.

