On October 29, 2026, Cursor’s extension marketplace served a tampered Claude Code update for eleven hours, from 2:27 AM to 1:35 PM Pacific. Up to 5,000 users potentially received versions 2.1.312 and 2.1.313. The update was the attack. Updating to the fixed version does not remove what it installed.
The 53-second reel: eleven hours, up to five thousand users, one backdoor.
The Window
The window opened at 2:27 AM Pacific on October 29 and closed at 1:35 PM the same day. Eleven hours. Cursor’s marketplace served the tampered builds from its own cache, so the update arrived through the channel users are taught to trust. At most 5,000 users who requested the specific cached versions potentially received the malicious build. That is the source’s exact qualifier: potentially received, not confirmed compromised. There is no confirmed count of successful workstation backdoors, and this post does not invent one. Cursor removed the malicious build after about eleven hours. The window closed. What entered through it did not leave with it.
The Payload
Two tampered builds, 2.1.312 and 2.1.313, carried the same payload. It added an attacker SSH key labeled svc-telemetry to the machine’s authorized_keys file, enabled the local SSH service, and attempted to open a reverse SSH tunnel to 134.209.82.225. It dropped persistence scripts that retry roughly every sixty seconds, and it collected host details: hostname, username, operating system, architecture, home directory, environment variable names, working directories, Git email, and public IP.
Two facts keep the payload honest. The key addition succeeded. The tunnel step usually failed. Turning on the local SSH service needed admin rights the extension usually lacked, and errors in the malicious code meant the tunnel would fail on most hosts. The payload is an attempted tunnel, not an accomplished one. The persistence is the part that matters: standalone scripts that survive the extension being uninstalled.
The Containment
The last attacker activity came November 1, and it was blocked. The attacker server was taken offline. The malicious build was pulled from the marketplace on October 29. Containment is not cleanup. Removing the extension, or updating past the tampered versions, does not remove the key from authorized_keys, does not remove the persistence scripts, and does not rotate anything the payload touched. That is the whole point of the check below: the fix lives on your machine, not in the vendor’s postmortem.
The Root Cause
On November 13, Cursor published the root cause. A key Cursor had issued to Fireworks AI, a vendor, was exposed by Fireworks. The attacker used the exposed credentials to reach Cursor’s systems and copied coding-agent traces from a shared read-only bucket. Less than 0.01 percent of customers had API credentials exposed. Traces were exposed for less than 0.05 percent of users. November 13 is the disclosure date. The key exposure itself happened before October 29.
Two keys, two different things. The leaked vendor key opened the door into Cursor’s systems. The attacker key, svc-telemetry, is the one the payload planted on your machine. Do not conflate them. The fix checks for the second one.
The Check
Run this tonight. It is our recommendation built from the source’s remediation steps, not the source’s verbatim text.
1. Check for the key. On macOS and Linux, run the source’s check:
grep -qs svc-telemetry ~/.ssh/authorized_keys && echo "FOUND" || echo "clean"
On Windows, check both authorized_keys paths in PowerShell:
Select-String -Path "$env:USERPROFILE\.ssh\authorized_keys" -Pattern "svc-telemetry"
Select-String -Path "C:\ProgramData\ssh\administrators_authorized_keys" -Pattern "svc-telemetry"
If the key is there, remove that line from the file.
2. Remove the persistence. On macOS, look for an unfamiliar LaunchAgent plist you did not install and remove it. On Windows, remove the SysMon scheduled task the payload created.
3. Rotate credentials. Rotate any credentials on the machine, starting with the ones the payload could read: SSH keys, API keys, and tokens stored in the home directory or environment.
Updating alone is not enough. The fixed version closes the window. It does not clean the machine.
A note on the fixed version, carried here only because the primary sources do not name it: version 2.1.314 is the fixed version according to secondary coverage only (Microsoft analysis via clearsecurity.vision, November 4; Phorion; BleepingComputer; unixshadow, November 2). Neither Cursor post mentions 2.1.314. It is not cited to them, and it does not appear in the reel.
Corroboration
The same month, Parameter Labs disclosed a different break in the same class of trust problem: a link you trust carrying something else. A full ChatGPT account takeover via a shared link took about 4.5 seconds from link to JavaScript running as chatgpt.com, reading private conversations. It was reported October 18, fixed October 26, with a $5,000 bounty. The post is dated October 28, 2026.
The pattern is the lesson. The update channel and the shared link are both trust infrastructure. When either one is poisoned, the fix is not to trust harder. It is to check the machine.
Sources
- Cursor, “Claude Code Extension Security Notice,” October 29, 2026, updated October 30, 11:30 AM PT: https://cursor.com/blog/claude-code-extension-security-notice
- Cursor, “Security Incident Update,” November 13, 2026: https://cursor.com/blog/security-incident-update
- Parameter Labs, “We Found a Gap at OpenAI,” October 28, 2026: https://www.parameter.ai/labs/we-found-a-gap-at-openai
This post started as an Instagram post →
- https://www.instagram.com/reel/Dfx0sfGCWuP/
- https://www.threads.com/@dataeng.ai/post/Dfx5f8SJZjP
- https://www.youtube.com/shorts/Dm9T8decymY
- https://cursor.com/blog/claude-code-extension-security-notice
- https://cursor.com/blog/security-incident-update
- https://www.parameter.ai/labs/we-found-a-gap-at-openai
Numbers above trace to these sources. If one moved, tell us and we fix it.


Talk it through
Argue with us on Instagram.