Cursor says an attacker used an access key exposed by AI infrastructure provider Fireworks to steal historical coding-agent data, reach internal services and distribute malicious copies of Anthropic’s Claude Code extension. The tampered builds were available through Cursor’s VS Code extension marketplace cache for roughly 11 hours on October 29, 2026, potentially reaching at most 5,000 users.
In its November 13 incident update, Cursor explicitly separates this attack from Anthropic’s official software and the Open VSX registry. Anthropic did not publish the malicious copies, its product was not the vulnerability, and Open VSX never hosted the malware, according to Cursor.
The immediate action is specific: anyone who installed or updated Anthropic.claude-code versions 2.1.312 or 2.1.313 through Cursor’s VS Code extension marketplace during the exposure window on Windows or macOS should check their system. Cursor warns that removing the extension may leave separate persistence scripts behind.
Who Should Check Their System
Cursor’s security notice identifies the following exposure conditions:
| Detail | Cursor’s reported scope |
|---|---|
| Extension identifier | Anthropic.claude-code |
| Affected versions | 2.1.312 and 2.1.313 |
| Platforms covered by the notice | Windows and macOS |
| Distribution path | Cursor’s VS Code extension marketplace cache |
| Exposure window | October 29, 2026, approximately 2:27 a.m. to 1:35 p.m. PT |
| Potential recipients | At most 5,000 users who requested the specific cached versions |
| Required response | Check for compromise and follow Cursor’s remediation guidance |
The 5,000 figure is a ceiling on potential recipients, not a confirmed count of infected computers or working backdoors. The incident details and exposure estimates come from Cursor’s investigation; the cited sources do not provide an independent forensic count. Fireworks’ separate bulletin also does not independently verify every step of Cursor’s account.
Cursor reports two additional data-exposure figures: fewer than 0.01% of customers had API credentials exposed, and historical coding-agent traces from non-Privacy Mode sessions were exposed for fewer than 0.05% of users. It says it contacted affected customers directly. These percentages describe different categories and cannot be added together or used to estimate how many people received the malicious extension. Source: Cursor’s incident update
I’d treat the extension window as its own check, separate from any data-exposure notification. Cursor says users who received no trace-exposure notification were not represented in the exposed traces. That assurance concerns historical data; it does not answer whether a machine downloaded a tampered extension.
The same distinction applies to Privacy Mode. Cursor says Privacy Mode prompts and code were not affected by the trace exposure. The extension warning, however, is defined by version, platform, distribution channel and installation time, with no Privacy Mode exemption.
How a Read-Only Key Became a Supply-Chain Attack
Cursor describes a chain of credential exposures crossing several systems.
First, the attacker obtained an access key Cursor had issued to Fireworks. According to Cursor, Fireworks exposed that key, which allowed the attacker to copy a read-only storage bucket used to share data with the vendor.
The bucket contained limited historical coding-agent traces, mostly from 2025, from sessions outside Privacy Mode that Cursor was permitted to use for model training. A small number contained credentials that had not been redacted. The attacker then used credentials from those traces to access selected software-as-a-service tools used by Cursor. Source: Cursor’s incident update
Cursor lists the information reached through those tools as:
- Archived, non-production Cursor source repositories, some containing business contact information.
- Limited customer-support data and email conversations.
- Product usage metrics and contact details.
- Summaries of some coding-agent conversations.
- Credentials users had sent to Cursor’s support team.
A stolen credential from exposed internal-system data then allowed the attacker to place malicious Claude Code packages directly into the cache serving Cursor’s VS Code extension marketplace.
The read-only permission did not contain the downstream risk. It limited what the attacker could do to that storage bucket, but the copied contents included credentials usable elsewhere. Those credentials supplied separate permissions, eventually including access sufficient to alter the extension distribution path.
That is the consequential failure illustrated by Cursor’s account: historical AI data can become an access route if it retains live secrets. Calling a dataset “offline” or a bucket “read-only” describes its intended use and storage permissions, not the authority carried by credentials inside it.
The cache detail also matters. A legitimate upstream registry can remain untouched while a downstream delivery system serves an altered package. Cursor’s account places the malicious copies in its own marketplace cache, explaining why the Claude Code name appeared in the incident without establishing a breach of Anthropic or Open VSX.
What the Malicious Extension Attempted
Cursor’s payload analysis describes information collection and an attempted persistent SSH backdoor.
The extension collected host and user details, including hostname, username, operating system, architecture, home-directory path, working directories, configured Git email and public IP address. It also collected environment-variable names. The notice does not describe that collection step as harvesting every environment-variable value.
The payload then attempted to:
- Add an attacker-controlled SSH public key, labeled
svc-telemetry, to authorized-key files. - Enable the local SSH service.
- Open a reverse SSH tunnel.
- Keep the tunnel running through a macOS LaunchAgent or Windows scheduled task, with a Windows startup-script fallback.
The background scripts retried the tunnel roughly every 60 seconds. A reverse tunnel creates an outbound connection that can provide a route back toward the machine, while the added SSH key supplies an authorization mechanism.
Cursor says administrative-permission requirements and errors in the malicious code prevented the SSH-enabling step from succeeding on most ordinary Windows and macOS hosts. It also says the extension added the attacker’s key, installed persistence and sent host details. Those are separate effects, so a failed remote-login setup is not equivalent to a clean machine.
Cursor reports that the attacker’s server was taken offline with help from its hosting provider, disabling the described reverse shell and host-detail submission. Server takedown does not remove files, keys or scheduled execution from a recipient’s computer. That is why uninstalling the extension alone may be insufficient.
Checks and Cleanup for Affected Users
Cursor provides platform-specific checks, remediation scripts and indicators in its current security notice. Use that page for the latest instructions and downloads, since the investigation’s findings may change.
For macOS, Cursor provides this check for the attacker’s key marker:
grep -qs svc-telemetry ~/.ssh/authorized_keys && echo "svc-telemetry key found" || echo "No key found; machine not impacted"
This command searches the current user’s SSH authorized-key file. It does not remove anything. Finding the marker is a reason to follow the remediation steps. Although Cursor’s command prints “machine not impacted” when it finds no match, the command itself checks one marker in one file; it is not a complete endpoint investigation.
For Windows, Cursor’s PowerShell check searches for svc-telemetry in these locations, where the files exist:
$HOME\.ssh\authorized_keysC:\ProgramData\ssh\administrators_authorized_keys
Administrators should also search endpoint, network and extension-installation telemetry using the notice’s indicators. The published indicators include modified package hashes, the svc-telemetry marker and the attacker address 134[.]209[.]82[.]225.
Cursor’s response guidance is to:
- Isolate hosts showing evidence of payload execution or backdoor installation.
- Remove the unauthorized SSH key, persistence mechanisms and running tunnel processes.
- Assess whether an affected host should be rebuilt.
- Rotate credentials and sessions accessible from affected hosts, using a trusted system.
Cursor says its remediation scripts remove the malicious extension, persistence mechanisms, tunnel processes and attacker key. They return a non-zero exit status if anything remains.
For macOS, download remediate-macos.sh from the notice and run it from its saved folder as the affected user, not as root:
bash remediate-macos.sh
For Windows, download remediate-windows.ps1, open PowerShell with Run as Administrator, and run it from its saved folder:
powershell -ExecutionPolicy Bypass -File .\remediate-windows.ps1
Frequently Asked Questions
4 questions
1Which Claude Code Versions Should Users Check?
Check Anthropic.claude-code versions 2.1.312 and 2.1.313 installed or updated through Cursor’s VS Code extension marketplace on Windows or macOS during the October 29 exposure window.