Security Digest · Nov 13, 2026 · Breach · AI agent tooling

Cursor Breach: Fireworks Key Led to Backdoored Extension

Cursor says a Fireworks-exposed key let attackers take agent traces and credentials, then ship a backdoored Claude Code extension to up to 5,000 users.

Our verdict

Check exposure

Most Cursor users are untouched, but anyone who pulled Claude Code 2.1.312 or 2.1.313 from Cursor's marketplace on October 29, or who got a credential notice from Cursor, has real cleanup to do; uninstalling the extension is not enough.

Who must act: Cursor users who installed or updated the Claude Code extension from the Cursor marketplace on October 29, 2026 (about 2:27am to 1:35pm PT), and customers or team admins Cursor notified about exposed keys or support data.

Who can relax: Users of Anthropic's official Claude Code extension from other marketplaces or Open VSX, Cursor users with Privacy Mode enabled who received no notice, and anyone with no svc-telemetry key on their machines.

Affected
Anthropic.claude-code 2.1.312 and 2.1.313 served by the Cursor marketplace on Oct 29, 2026 (macOS and Windows builds)
Cursor coding-agent traces (mostly 2025) from sessions without Privacy Mode, shared with Fireworks
Credentials customers sent to Cursor support
Archived non-production Cursor source repositories, support emails, usage metrics and contact details
Fixed in
Malicious versions removed from the Cursor marketplace on Oct 29, 2026
Cursor marketplace rebuilt; Fireworks key revoked and inference moved off Fireworks
Cursor remediation scripts remediate-macos.sh and remediate-windows.ps1
CVSS
Not scored
EPSS
n/a
CISA KEV
Not listed
Exploitation
Exploited in the wild
Severity
High

What to do

  1. Check every developer machine for the svc-telemetry key in ~/.ssh/authorized_keys (and C:\ProgramData\ssh\administrators_authorized_keys on Windows).
  2. On any hit, isolate the machine and run Cursor's remediation script, or remove the key, persistence and tunnel processes by hand; uninstalling the extension does not remove them.
  3. From a clean machine, rotate every credential and session the affected user could reach: cloud, Git, SSH, registry, AI provider and SaaS tokens.
  4. Rotate every key Cursor listed in its notification email to team admins, even if you see no sign of use.
  5. Rotate anything you have ever pasted into a Cursor support ticket.
  6. Turn on Privacy Mode for teams that handle secrets, and keep secrets out of agent sessions.

How to check if you are affected or compromised

  • grep svc-telemetry ~/.ssh/authorized_keys, or look for key fingerprint SHA256:0Z6R+DF64Do2nwQic8a7kkfDLNNk4IiK+2U2Rci4htk.
  • Look for ~/.config/.smt, ~/.config/.smtk, $TMPDIR/.stk_*, ~/Library/LaunchAgents/com.apple.sysmond.plist on macOS, and a SysMon scheduled task or sysmon.cmd Startup script on Windows.
  • Search EDR and network logs for connections to 134.209.82[.]225 on TCP 22 or 8099, and for ssh or autossh running with -R *:0:localhost:22.
  • Check extension directories for anthropic.claude-code-2.1.312-* or -2.1.313-*, or the extension.js SHA-256 ecb58cb29078d9646ad1dcdbe51526f2d9cfea06be5fb2ab183e8ecc3ab6bf15.

What happened#

On November 13, 2026, Cursor published an incident update explaining how a trojanized Claude Code extension ended up in its marketplace on October 29. The root cause was a third party. An attacker obtained an access key that Cursor had issued to Fireworks AI, its model training and inference provider, and that Fireworks then exposed. With it, the attacker copied a read-only storage bucket Cursor used to share data with Fireworks. The bucket held coding-agent traces, mostly from 2025, from sessions without Privacy Mode. Some traces contained unredacted credentials.

The attacker used those credentials to get into several Cursor SaaS tools. From there they took archived non-production source repositories, limited support data and emails, usage metrics and contact details, summaries of some agent conversations, and credentials customers had sent to Cursor support. A stolen credential then let them place two builds, labeled Anthropic.claude-code 2.1.312 and 2.1.313, on the cache that serves Cursor's VS Code extension marketplace. Cursor says Anthropic did not develop these builds. They were live for about 11 hours, from roughly 2:27am to 1:35pm PT on October 29, and Cursor estimates up to 5,000 users requested them.

According to Cursor's notice and Phorion's analysis, the payload was a loader prepended to extension.js that ran as soon as the extension loaded. It collected the hostname, username, OS, home and working directories, the global Git email, the public IP and up to 40 environment variable names matching KEY, TOKEN, SECRET, PASS, AWS or GCP. It did not collect their values. It added an SSH key with the comment svc-telemetry to authorized_keys, tried to enable the local SSH server, and opened a reverse SSH tunnel to 134.209.82.225. It made this persistent through a com.apple.sysmond LaunchAgent on macOS, or a SysMon scheduled task with a Startup-folder fallback on Windows, retrying about every 60 seconds.

Who is affected#

There are two groups. The first is anyone who installed or auto-updated the Claude Code extension from Cursor's marketplace during the October 29 window. Cursor says the official Anthropic extension, Open VSX, other marketplace extensions, the Cursor app, its updates and its signing keys were not affected. The second is customers whose data was in the exposed traces or SaaS tools. Cursor puts exposed API credentials at fewer than 0.01% of customers and exposed traces at fewer than 0.05% of users. It says it has notified both groups by email and in-app notices, and that users who received no notification were not in the exposed traces. Privacy Mode sessions were not included.

Do you really need to act?#

Only if you are in one of those two groups, but then fully. Phorion found bugs that limited the backdoor. On macOS, the Remote Login check runs without admin rights and fails, so SSH is not switched on. On Windows, the OpenSSH setup calls /bin/sh, which normal installs lack. On a laptop where SSH was not already enabled, the tunnel most likely gave the attacker no shell. However, the key and persistence were still installed, and the reconnaissance data was still sent. Machines that already ran an SSH server, which is common on developer Macs and Linux boxes, could have been reached. The attacker server is now offline, which does not undo access that may already have happened.

For notified customers, the risk is that leaked keys get used. Cursor has emailed admins lists of keys and asked third-party services to revoke them. Do not rely on that: rotate the keys yourself.

What to do now#

  1. On macOS or Linux, run grep svc-telemetry ~/.ssh/authorized_keys. On Windows, run Select-String svc-telemetry $env:USERPROFILE\.ssh\authorized_keys, C:\ProgramData\ssh\administrators_authorized_keys. Because the comment can be edited, also compare key fingerprints with ssh-keygen -lf ~/.ssh/authorized_keys.
  2. If you find it, isolate the machine and run Cursor's script (bash remediate-macos.sh as the user, or remediate-windows.ps1 from an Administrator PowerShell). Cursor says the script exits non-zero if anything remains. Consider rebuilding the machine.
  3. From a clean machine, rotate everything the user could reach, including Git and SSH keys, cloud credentials, package registry tokens and AI provider keys. Revoke sessions too.
  4. Rotate every key on the list Cursor sent your admins, and anything you have ever pasted into a Cursor support thread.
  5. Enable Privacy Mode for teams that work with secrets.

How to tell if you were hit#

Look for ~/.config/.smt, ~/.config/.smtk and $TMPDIR/.stk_*. On macOS, check for ~/Library/LaunchAgents/com.apple.sysmond.plist with launchctl list | grep sysmond. On Linux, look for a crontab -l entry with @reboot running .smt. On Windows, run schtasks /query /tn SysMon and check for sysmon.cmd in the Startup folder. In network and EDR data, look for connections to 134.209.82[.]225 on ports 22 or 8099, ssh or autossh with -R *:0:localhost:22, and IDE helper processes writing to authorized_keys. A successful SSH login from 127.0.0.1 with the planted key means an operator came in through the tunnel. The backdoored extension.js SHA-256 is ecb58cb29078d9646ad1dcdbe51526f2d9cfea06be5fb2ab183e8ecc3ab6bf15.

Timeline#

  • — Malicious 2.1.312 and 2.1.313 builds are served from Cursor's marketplace for about 11 hours, and Cursor posts its first notice.
  • — Cursor publishes IOCs and remediation scripts and begins notifying users. BleepingComputer reports the SSH backdoor.
  • — Last attacker activity Cursor observed, which was blocked.
  • — Cursor discloses the Fireworks key as the root cause and the wider data exposure.

Sources

Original sources first: vendor and project advisories, CVE and government records, the discoverers’ write-ups. Every claim above links to one of these. See how the Security Desk works.

Original and primary sources

  1. Cursor — Security incident update (Nov 13, 2026) cursor.com
  2. Cursor — Security notice: Anthropic.claude-code 2.1.312 and 2.1.313 (Oct 29, updated Oct 30, 2026) cursor.com
  3. Phorion — Official Claude Code VSX compromise phorion.io

Coverage and analysis

  1. BleepingComputer — Cursor confirms security breach led to malicious Claude Code extension bleepingcomputer.com
  2. BleepingComputer — Malicious Claude Code extension on Cursor planted SSH backdoors bleepingcomputer.com