# How a Leaked Key Led to Malicious Claude Code Extensions in Cursor

How a leaked Fireworks key led to tampered Claude Code extensions in Cursor, which installs were at risk, and how to check a Mac or Windows PC.

Canonical URL: https://aiidelist.com/blog/cursor-claude-code-extension-supply-chain-attack

Language: en

Published: 2026-11-15

Updated: 2026-11-15

## Key Takeaways

- **Check where and when the extension was installed.** On October 29, 2026, Cursor’s VS Code extension marketplace served tampered copies of `anthropic.claude-code` versions `2.1.312` and `2.1.313` for macOS and Windows. The exposure window was approximately **09:27–20:35 UTC**. [Cursor’s security notice](https://cursor.com/en-US/blog/claude-code-extension-security-notice)
- **Up to 5,000 users may have received a malicious package.** That figure counts potential recipients. Establishing whether the code ran or someone logged in requires evidence from the individual workstation. [Cursor’s November 13 investigation](https://cursor.com/blog/security-incident-update)
- **Two groups need to act for different reasons.** Developers who installed an affected build should check their machines. Customers whom Cursor notified about exposed API credentials should rotate the identified keys, even if they never installed Claude Code. [Cursor’s incident report](https://cursor.com/blog/security-incident-update)

## How a read-only key led to a software supply chain attack

The incident began with a key that Cursor had issued to Fireworks. It granted read-only access to a storage bucket containing historical coding-agent traces, mostly from 2025. Some traces still contained working credentials. The storage permission prevented changes to the bucket, but anyone who copied a credential could use whatever permissions it carried elsewhere. [Cursor’s investigation](https://cursor.com/blog/security-incident-update)

According to Cursor, credentials in the exposed data led the attacker into internal services. A further stolen credential let the attacker place altered extension packages directly in the cache that served Cursor users. A developer could therefore select the familiar Claude Code listing and receive different software from someone downloading the same version through another channel. [Cursor’s account of the marketplace compromise](https://cursor.com/blog/security-incident-update)

That detail explains why the **download channel** matters. Cursor says Open VSX, its upstream registry, never hosted the malicious packages. Microsoft confirmed that neither its Visual Studio Marketplace nor the official Claude Code npm package distributed them. A developer who used only the npm CLI falls outside the reported extension exposure. A developer who also used Claude Code inside Cursor should check that installation separately. [Microsoft Security Research’s analysis](https://www.microsoft.com/en-us/security/blog/2026/10/30/inside-malicious-claude-code-branded-cursor-extension-persistent-ssh-backdoor/)

## Who was affected by the Cursor breach?

Cursor reported three figures with different denominators. They answer different questions and should never be added together. [Cursor’s November 13 incident report](https://cursor.com/blog/security-incident-update)

| Reported exposure | What the figure measures | What to check |
|---|---|---|
| Up to 5,000 users | People who requested specific cached extension builds and may have received a tampered copy | Workstation install and update history, package hashes, and evidence of execution |
| Fewer than 0.05% of users | Users whose historical coding-agent traces were exposed | Cursor’s direct notification about the affected traces |
| Fewer than 0.01% of customers | Customers whose API credentials were exposed | The specific keys Cursor identified to the team administrator and their usage logs |

**Privacy Mode matters to the second row.** Cursor says prompts and code from Privacy Mode sessions were outside the exposed trace data. A developer could still have downloaded an affected extension while using Privacy Mode. Conversely, a team could have a credential to rotate without finding a single affected workstation. [Cursor’s findings on Privacy Mode and customer notifications](https://cursor.com/blog/security-incident-update)

## How to check a Mac or Windows PC

Start with the installation record for **October 29, 02:27–13:35 Pacific daylight time**. Include automatic updates. The version displayed in Cursor today tells you what is installed now. An extension update during or after the incident can make that screen a poor record of what ran that day. Cursor published separate SHA-256 hashes for each affected version and platform, covering Apple silicon Macs (`darwin-arm64`) and 64-bit Windows (`win32-x64`). [Cursor’s package hashes and exposure window](https://cursor.com/en-US/blog/claude-code-extension-security-notice)

**If the original `.vsix` package is available**, calculate its hash without opening or installing it. Replace `extension.vsix` with the path to the retained download.

On macOS:

```bash
shasum -a 256 "extension.vsix"
```

On Windows PowerShell:

```powershell
Get-FileHash -Algorithm SHA256 -LiteralPath 'extension.vsix'
```

Compare the result with Cursor’s entry for that **exact version and platform**. A match identifies a malicious package and warrants checking whether it loaded. Hashing a fresh download only identifies the file retrieved today. If the original package is gone, use extension-installation records and endpoint logs. [Cursor’s indicators of compromise](https://cursor.com/en-US/blog/claude-code-extension-security-notice)

**If the extension may have loaded**, check for the attacker’s SSH public key. On a Mac, this read-only command searches the current user’s authorized keys for the published marker:

```bash
grep -nF 'svc-telemetry' "$HOME/.ssh/authorized_keys"
```

Compare any match with the **full key fingerprint** in Cursor’s notice. An empty result only describes that file for that user. On Windows, Cursor’s check covers both the user’s authorized keys and the system-wide administrators’ key file. The notice also identifies startup artifacts on both platforms. [Cursor’s platform-specific checks](https://cursor.com/en-US/blog/claude-code-extension-security-notice)

## What the malicious extension could do

Microsoft found that the injected JavaScript ran as soon as the extension module loaded, before normal initialization. The developer did not need to send a Claude prompt. The code collected host details, attempted to establish SSH access, and wrote startup files that could survive removal of the extension. Its initial reporting code collected selected environment-variable **names**. That detail matters when investigating a suspected credential theft: seeing a variable name in the report does not establish that its secret value was transmitted. [Microsoft’s technical analysis](https://www.microsoft.com/en-us/security/blog/2026/10/30/inside-malicious-claude-code-branded-cursor-extension-persistent-ssh-backdoor/)

A failed SSH connection also leaves an open question. Cursor reported that implementation errors and limited operating-system privileges prevented the backdoor from working on many machines. An unauthorized key or startup file could still remain. To determine whether anyone actually accessed a workstation, review SSH authentication records and subsequent process and file activity alongside the installation timestamp. [Microsoft’s investigation and recovery guidance](https://www.microsoft.com/en-us/security/blog/2026/10/30/inside-malicious-claude-code-branded-cursor-extension-persistent-ssh-backdoor/)

## What to do if a check finds a match

1. **Isolate a workstation that shows evidence of execution.** On a managed device, involve the security team and preserve the original package, timestamps, and endpoint logs before cleanup. These records help distinguish a downloaded extension from a successful login. [Microsoft’s response guidance](https://www.microsoft.com/en-us/security/blog/2026/10/30/inside-malicious-claude-code-branded-cursor-extension-persistent-ssh-backdoor/)
2. **Run the official cleanup for the affected operating system.** Cursor instructs Mac users to run its script under the affected user account. Its Windows script requires an administrator PowerShell session. Both scripts report a nonzero exit status if artifacts remain. [Cursor’s remediation scripts and instructions](https://cursor.com/en-US/blog/claude-code-extension-security-notice)
3. **Recover credentials from a trusted device.** Rotate credentials accessible from the affected workstation and revoke its sessions. Separately, if Cursor identified a specific API key to a team administrator, replace that key, revoke the old one, and review the provider’s usage history for unexpected activity. [Cursor’s customer notification and rotation guidance](https://cursor.com/blog/security-incident-update)
4. **Verify recovery after login or restart.** Check that the unauthorized key and startup items stay gone. If the investigation cannot establish trust in the workstation, assess rebuilding it before returning it to development work. [Microsoft’s recovery criteria](https://www.microsoft.com/en-us/security/blog/2026/10/30/inside-malicious-claude-code-branded-cursor-extension-persistent-ssh-backdoor/)

## Conclusion

The most useful record in this incident is the one that connects **a specific workstation, download channel, version, and time**. Check October 29 installation history against Cursor’s published window. If a device matches, follow the endpoint checks through recovery. If the team received a credential notice, assign an owner to each named key and confirm that its old value has been revoked. Those two checks address the two distinct ways this incident reached developers. [Cursor’s incident report](https://cursor.com/blog/security-incident-update) · [Official workstation checks](https://cursor.com/en-US/blog/claude-code-extension-security-notice)
