You install a new AI code editor. It offers to import your extensions and settings from VS Code. You click yes. Everything looks exactly like yesterday.
That click is where a supply chain risk called extension confusion begins. On a recent episode of Scale To Zero, I talked with Raj Handa, Senior Cybersecurity Engineer in offensive security at Workday, where he leads the pentest program and runs bug bounty. Raj introduced extension confusion at the Bay Area OWASP August meetup.
Here are the practical parts of that conversation, for anyone who owns developer tooling, endpoint security, or AppSec.
What Is Extension Confusion?
Modern developer tooling is split across two extension ecosystems.
VS Code pulls extensions from Microsoft’s Visual Studio Marketplace, whose terms, as Raj explained, restrict it to official Microsoft products. So AI-first editors such as Cursor and Windsurf typically turn to Open VSX, an open source registry run by the Eclipse Foundation.
Both registries use the same naming scheme: publisher.extension. But they do not share accounts, namespaces, or ownership.
“A name on one registry has zero legal or, I would say, cryptographic link to the same name on the other.” (Raj Handa)
Here is how the attack plays out, as Raj described it:
-
A developer installs a new AI editor and imports their extensions from VS Code.
-
The editor copies the local extensions but points future updates at Open VSX.
-
If one of those extensions was never registered on Open VSX, its namespace there is orphaned. An attacker claims the name and publishes a higher version.
-
Auto-updates are on by default, so the editor quietly downloads the attacker’s version in the background.
“Zero clicks, no prompt, and no warning.”
Extensions are not sandboxed. They run in your developer context and inherit your shell, SSH keys, secrets, and cloud tokens. Raj said a benign test he ran against unclaimed namespaces kept getting callbacks from managed corporate laptops for over 60 days.
Open VSX’s namespace documentation notes that extensions in namespaces without an owner are shown as unverified with a warning icon. That helps a human who looks. Fewer and fewer installs involve a human looking.
How Is It Different From Dependency Confusion?
Dependency confusion tricks the build: a package manager looking for an internal library queries PyPI or npm by mistake and pulls a higher public version.
“Extension confusion moves that trick up to the editor level.”
Dependency confusion is an internal versus public problem. Extension confusion, as Raj put it, is an ecosystem fracture between two parallel public registries: the attacker takes a trusted public tool from one store and exploits the fact that another public store lets anyone claim that identity.
Doesn’t Registry Scanning Solve This?
Raj’s answer was “yes and no.” Scanning is a good first step, but he noted that attackers have already slipped credential-stealing payloads hidden with invisible Unicode past Open VSX scanning. You still need complementary controls.
Why the Editor Is a Security Blind Spot
We scan libraries, sign commits, and harden CI/CD. Raj’s view is that the problem is where we drew the trust boundary.
“We spent decades hardening the factory, but we left the workbench wide open.”
The editor sits upstream of every pipeline control, on a laptop that holds keys to production. Raj gave three reasons this is so dangerous:
-
Zero sandboxing. An extension runs with the developer’s full permissions: environment files, tokens, uncommitted code, cloud credentials, SSH keys.
-
Detection asymmetry. Developer laptops are noisy, so odd outbound traffic gets written off as normal activity.
-
Workstations treated as scratch pads. We assume risk starts at merge. With AI-assisted IDEs, Raj argued, the workstation is effectively production-like.
AI Agents and Devcontainers Remove the Human
Raj thinks this is getting worse. Traditionally an attacker still needed a human to ignore an unverified publisher and click install.
Now you clone a repo, click “trust folder,” and the editor installs whatever devcontainer.json or extensions.json lists, resolved strictly by name. Nobody checks the publisher badge. Coding agents go further, deciding they need a linter or formatter and installing it by name on the fly.
“When you bring autonomous coding agents into your environment, you do not really apply any human skepticism.”
Real-World Signal: 77 Evil Twins on Open VSX
In August 2026, Manifold Security reported 77 counterfeit “evil twin” extensions on Open VSX, each copying the name, namespace, and description of a real VS Code Marketplace extension and beaconing to the same new domain.
Nineteen sent details like the git remote, commit email domain, workspace path, branch, and CI project identifiers, described in the listings under “Telemetry.” Manifold says the packages were removed from Open VSX as of August 3, 2026. Raj’s read:
“An extension beaconing out with internal Git metadata doesn’t really look like an intrusion. It just looks like developer noise.”
What Security Teams Should Lock Down
Raj gave three core controls. The last two items follow from the rest of our conversation.
Extension Security Checklist
-
Allowlist by verified publisher ID, not by name. “If your policy is only checking the name string, you have secured nothing,” Raj said. Install only when both the approved extension ID and the verified publisher match.
-
Resolve from an internal mirror. Treat extensions like enterprise Python or npm packages: mirror internally, curate what enters, and point editors only at that source.
-
Turn off silent auto-updates and pin versions. Counterintuitive, but the risk is an extension that is benign on Monday and updates itself into a malicious payload on Tuesday morning. Pinning buys time to detect it.
-
Audit repo-driven installs. Review what devcontainer.json and extensions.json ask editors and agents to install.
-
Keep secrets off the workstation. Fewer long-lived credentials on laptops means a smaller blast radius.
For VS Code, Microsoft documents an extensions.allowed setting and AllowedExtensions policy that allow or block by publisher, extension ID, and version. Check what your AI editor offers before rollout.
Who Owns This?
Extension confusion slips through partly because nobody clearly owns the editor. AppSec owns the pipeline, endpoint teams own the laptop, and developers own their setup.
Raj’s framing points to an answer: platform teams should own extensions the way they own package registries. One practical split: platform or developer-experience teams run the mirror and editor configuration, security sets the allowlist criteria and detections, and a new AI editor gets reviewed like any third-party tool before rollout.
One Bad Habit to Break
Asked by a listener for the one habit he would stop every developer from doing, Raj said: stop hard-coding an API key to test locally and promising to clean it up before you push. Extensions inherit everything on that machine.
“In software, nothing is more permanent than a temporary workaround.”
For an offensive team, those forgotten shortcuts often mean there is no need to burn a zero-day at all.
Listen to the Full Episode
The full episode also covers running a bug bounty when much of the inbox is AI-generated noise, and pentesting the third-party supply chain. On that last topic, see Raj’s Workday engineering post, Beyond the Vendor Perimeter: Strategic Pentest for Third-Party Supply Chain Attacks.
-
Watch on YouTube: Pentesting, Bug Bounties & The IDE Supply Chain Blind Spot ft. Raj Handa
-
Listen on Apple Podcasts or Spotify
-
Join the conversation on LinkedIn
Scale To Zero is the security podcast we run at Cloudanix. If this was useful, subscribe and share it with whoever owns your developer tooling. And if your team has already locked down editor extensions, I would love to hear how.
People Also Read
- AI Slop in Bug Bounty: How to Keep Your Program From Becoming a Spam Folder
- How Often Should You Pentest? Treat Penetration Testing as a Portfolio
- Never Pay for the Same Bug Twice: Vulnerability Prioritization Beyond CVSS
- Bug Bounty and Pentesting Through M&A: How to Bring an Acquisition Into Scope
- A Vendor’s Exposed Key Put a Fake Claude Code Extension in Cursor’s Marketplace: What Happened and How to Contain It
- When the AI Coding Agent Becomes the Attacker: Lessons From the Cursor Breaches