Cursor has named the root cause of the Cursor malicious extension incident. On November 13, Cursor said it started with an access key it had issued to Fireworks AI. Cursor says Fireworks exposed that key. From there, the attacker worked through Cursor’s systems. On October 29, the attacker placed tampered copies of the Claude Code extension in Cursor’s marketplace. Cursor says at most 5,000 users could have received them.
Cursor is an AI code editor built on VS Code, used by many agencies and in-house dev teams. Our explainer on what Cursor is covers the basics. This post covers what Cursor now says about the Cursor malicious extension, what it did, and how to check your machines. It also covers what to rotate, Privacy Mode, and how to limit extensions. Facts are as of November 16, 2026.
The short version: did anyone on your team install or update the Claude Code extension in Cursor on October 29? If so, run Cursor’s check commands now. If the attacker’s SSH key is there, isolate that machine and rotate every credential it could reach. Also check that Privacy Mode is on for your team. Ask your vendors how they store the keys you issue them.
What happened in the Cursor malicious extension incident
Cursor published two posts. The first, its Claude Code extension security notice, went up on October 29 and was updated October 30. The second, "An update on our recent security incident", is dated November 13.
| Date (2026) | Event | Source |
|---|---|---|
| October 29, 2:27 a.m. to 1:35 p.m. PT | Tampered Claude Code extension versions 2.1.312 and 2.1.313 served from Cursor’s marketplace | Cursor notice, Oct 29 |
| October 30 | Notice updated with revised indicators and affected versions; affected users notified | Cursor, Oct 30 and Nov 13 |
| November 1 | Last attacker activity Cursor saw, which it blocked | Cursor update, Nov 13 |
| November 13 | Cursor names the root cause: an access key issued to Fireworks | Cursor update, Nov 13 |
Cursor’s notice names the extension as "Anthropic.claude-code version 2.1.312 and 2.1.313 for Windows and MacOS." It says these versions "were not developed by Anthropic." Cursor’s November 13 update adds that the marketplace source, Open VSX, never had the malware. It says the Cursor malicious extension was available for about 11 hours.
How a vendor’s key reached the marketplace
Cursor’s November 13 update lays out the chain. An attacker used an access key Cursor had issued to Fireworks, an AI inference provider. Cursor says Fireworks had exposed that key. With it, the attacker copied data from a read-only storage bucket Cursor used to share data with Fireworks.
That bucket held coding agent traces, mostly from 2025. Cursor says they came from sessions not in Privacy Mode, which Cursor was allowed to use for model training. Some traces held credentials that had not been redacted. According to Cursor, "less than 0.01% of customers had API credentials exposed" this way.
The attacker used those credentials to reach some of the SaaS tools Cursor uses. Cursor lists what was accessed:
- Archived non-production source code repositories.
- Limited customer support data and emails.
- Product usage metrics.
- Summaries of some agent conversations.
- Credentials users had sent to support.
Then, using a stolen credential, the attacker put tampered copies of the extension in the cache that serves Cursor’s marketplace. That is how the Cursor malicious extension reached users who requested those versions on October 29.
Fireworks published its own October 2026 security incident bulletin, last updated October 10. It says Fireworks found "unauthorized use of internal Fireworks credentials" on October 6. It says an outside party reached environment variables used in evaluation jobs in "a limited number of customer accounts" at Fireworks. The bulletin does not name Cursor. Cursor’s update does not say when its key was exposed. The Hacker News linked the two in its coverage on November 15. Neither company confirms that link on its own pages.
What the Cursor malicious extension did
Cursor’s October 29 notice describes the payload. According to Cursor, it:
- Sent the hostname, username, operating system, architecture and public IP address to an attacker server.
- Tried to add an attacker-controlled SSH key, turn on the SSH service and open a reverse SSH tunnel.
- Tried to keep that tunnel alive with a macOS LaunchAgent or a Windows scheduled task. It retried about every 60 seconds.
Cursor says the SSH steps would usually fail without administrator rights. It also says the attacker server is offline. But the notice warns that the limited data collection "does not rule out later credential access over the backdoor" that the payload tried to open. That is why Cursor tells affected users to rotate credentials, not just remove the extension.
What Cursor says was not affected
Cursor’s November 13 update lists what it found no impact on:
- Prompts and code for Privacy Mode users.
- Production systems and the infrastructure that runs Cursor.
- The Cursor app, official updates and shipped code.
- Other marketplace extensions and customer code repositories.
On the app, Cursor says it builds from a separate code host the attacker never reached, and "no release signing keys were exposed." For users not in Privacy Mode, Cursor says traces were exposed "for fewer than 0.05% of users," and each was notified directly.
How to check if you were hit
Start with who installed what. Ask each developer on your team if they use the Claude Code extension in Cursor. Then ask if it updated on October 29. In Cursor, the Extensions panel shows the installed version.
Then run Cursor’s check on each machine that might have had the Cursor malicious extension. Cursor gives this command for macOS:
grep -qs svc-telemetry ~/.ssh/authorized_keys && echo "svc-telemetry key found" || echo "No key found; machine not impacted"
On Windows, the notice gives a two-line PowerShell check. It looks in both your user authorized_keys file and C:\ProgramData\ssh\administrators_authorized_keys. Copy it from Cursor’s notice rather than retyping it.
Cursor also publishes remediation scripts, remediate-macos.sh and remediate-windows.ps1. The notice lists file hashes, an attacker IP address and process names too. If you have endpoint monitoring or an MSP, send them the notice. Ask them to search for those indicators.
Rotate SSH keys and secrets
If the svc-telemetry key turns up, Cursor’s notice lists the steps. Isolate the host. Remove the unauthorized SSH key, the persistence items and any running tunnel processes. Then decide whether to rebuild the machine. Last, rotate "credentials and sessions accessible from affected hosts" from a trusted system.
For an agency developer, that can be a long list:
- SSH keys used to reach client servers and Git hosts.
- Hosting panel and CMS admin sessions.
- Cloud provider credentials and AI provider API keys.
- Tokens stored in
.envfiles or editor settings.
Do the rotation from a clean machine, not the affected one. Our post on Claude Code settings files leaking secrets shows where tokens tend to hide in a project folder.
Cursor may also have told you that your credentials were in the exposed traces. If so, its update says to rotate the credentials listed for you. Cursor says it asked the issuing services to revoke them.
Check Privacy Mode for your team
The exposed traces came from sessions not in Privacy Mode. Cursor’s privacy help page says "Privacy Mode ensures your code is never used for training by Cursor or other AI model providers" when it is on. The page says Privacy Mode is on by default for team members and for Enterprise teams. The page does not give a default for individual plans.
To check, open Cursor Settings with Cmd+Shift+J on a Mac or Ctrl+Shift+J on Windows and Linux. Choose General and look for the Privacy Mode toggle. Freelancers and contractors on personal accounts are the ones to check first.
Cursor’s security hardening guide for admins goes further. It says enforcing Privacy Mode org-wide means members cannot turn it off. It also notes a gap. When people use their own API keys, data retention depends on their own agreement with that model provider.
Lock down which extensions your team installs
The same hardening guide lists two extension controls. Allowed Extensions lets admins allowlist trusted publishers. The guide says any entry blocks everything else unless you add a wildcard. An extension install cooldown holds back new versions until they have been public for a set number of hours. The guide says the goal is "to blunt short-lived malicious uploads."
A cooldown fits the Cursor malicious extension case closely. The tampered build was live for about 11 hours. The guide says most enforcement levers, such as org-wide policies and MDM, are Enterprise features. Smaller teams can still keep a written list of approved extensions. They can also turn off automatic extension updates on shared machines.
This is not the first warning about coding tools that trust what they are handed. See our posts on GitSpawn and repo config, Plugin4Shell and pinned plugins and the VS Code Workspace Trust bypass.
Vendor key hygiene: what this teaches agencies
The Cursor malicious extension started with a key given to a partner, not a flaw in the editor. Agencies issue keys like this all the time, to hosting companies, analytics tools, form builders and AI vendors.
Cursor’s list of fixes is a useful checklist for your own vendors. Cursor says it moved all inference off Fireworks and turned off the key it had issued. It cut off Fireworks’ access to shared datasets. It rotated every exposed credential it found and added secret scanning across code repositories. It also tightened network access for service providers. Apply the same ideas before an incident:
- Scope every key. Read-only, one bucket, one project.
- Set expiry dates. Our guide to OpenAI API key expiration shows one way to do it.
- Keep secrets out of logs and traces. Unredacted credentials in agent traces turned a data leak into a breach.
- Keep an inventory. Know which vendor holds which key, so you can turn it off in minutes.
The npm supply chain attacks on build servers showed the same pattern. Whatever holds your keys is part of your attack surface.
Frequently Asked Questions
What was the Cursor malicious extension?
Tampered copies of the Claude Code extension, versions 2.1.312 and 2.1.313 for Windows and macOS, served from Cursor’s marketplace on October 29, 2026. Cursor says Anthropic did not build them.
How did the attacker get in?
Cursor says an attacker used an access key it had issued to Fireworks AI, which Fireworks exposed. Credentials in copied coding agent traces then led to Cursor’s SaaS tools and a credential used on the marketplace cache.
How many people were affected?
Cursor says at most 5,000 users who requested specific cached versions on October 29 could have received the tampered build. It notified affected users directly.
How do I know if my machine was hit?
Run the check in Cursor’s October 29 notice. It looks for an SSH key with the comment svc-telemetry in your authorized_keys files. If it is there, the machine was affected.
Is uninstalling the extension enough?
No. Cursor says the payload drops persistence scripts that keep running. Follow the full removal steps and rotate credentials from a clean machine.
Were Privacy Mode users affected by the Cursor malicious extension incident?
The extension risk depended on installs, not Privacy Mode. For the trace leak, Cursor says prompts and code of Privacy Mode users were not affected.
Was the Cursor app itself compromised?
Cursor says no. It builds the app from a separate code host the attacker never reached, and no release signing keys were exposed.
What did Fireworks AI say?
Fireworks’ bulletin describes unauthorized use of internal credentials found on October 6, reaching environment variables in a limited number of customer accounts. It does not name Cursor.
How can admins limit extensions in Cursor?
Cursor’s hardening guide lists an Allowed Extensions setting for trusted publishers and an install cooldown that delays new versions. Most enforcement features are for Enterprise plans.
What should I ask vendors that hold my keys?
Ask how they store your keys, who can reach them, how fast they would tell you about a leak, and whether keys appear in logs. Scope and expire every key you issue.