Ghostdrive

News · Security · November 14, 2026

Cursor: a vendor key, agent traces and a tampered extension

Cursor says a vendor key led to agent traces holding credentials and a tampered Claude Code extension. What is stated, what is not, what to do.

DRIVE 369 DAO · 4 min read

Cursor says an exposed key at a vendor led to coding-agent traces that held credentials, and then to a tampered copy of the Claude Code extension on its marketplace. The company's security notice of October 29 described the extension; its update of November 13 gives the root cause. Everything below is Cursor's own account, read on November 14, 2026; we have no independent confirmation. Fireworks has published its own bulletins (October 9, updated November 9); they do not mention Cursor, and we have not compared the two accounts.

What Cursor states

QuestionCursor's accountTier
Where did it start?At Fireworks, a third-party training and inference vendor. An access key Cursor had issued to Fireworks, which Cursor says Fireworks exposed, was used to copy data from a read-only Cursor bucket shared with FireworksReported by the company
What was in the bucket?Coding-agent traces; some held unredacted credentialsReported by the company
What followed?The attacker used those credentials to reach SaaS tools, then a stolen credential to place unauthorised copies of Anthropic's Claude Code extension on the marketplace cacheReported by the company
Which versions?Anthropic.claude-code 2.1.312 and 2.1.313, Windows and macOS, about 2:27 am to 1:35 pm Pacific on October 29, about 11 hoursReported by the company
How many users?At most 5,000 users who requested those cached versions may have received itReported by the company
What did the malicious version do?Collected host details, added an attacker SSH key labelled svc-telemetry and persistence scripts. Cursor says turning on SSH needs admin rights and fails on most hosts, and that the attacker's server has been taken offlineReported by the company
Who built it?Not Anthropic; Cursor says this was not a vulnerability in Anthropic's productReported by the company
What data was touched?Archived non-production source repositories, limited support data, usage metrics, summaries of some agent conversations, credentials sent to support; API credentials for under 0.01% of customers; coding-agent traces for fewer than 0.05% of users (Cursor does not say of which group), for people not using Privacy ModeReported by the company

Cursor says Privacy Mode users' prompts and code were not affected and that no release signing keys were exposed.

What Cursor does not say

It does not name the attacker or a motive, say how Fireworks exposed the key, list which services the exposed credentials belonged to, or give an exact count of affected customers. Fireworks is also on the list of 16 AI platforms with disclosed incidents that Wiz compiled; whether the two cases are the same event is not stated by Cursor, Wiz or Fireworks.

Do this week

Steps 1 and 2 follow Cursor's notice; steps 3 to 5 are our general advice.

  1. If you use Cursor, check whether you installed or updated Claude Code to 2.1.312 or 2.1.313 on October 29 and follow Cursor's notice: look for the svc-telemetry SSH key, search for the listed indicators, isolate a host that shows them, run Cursor's remediation script or remove the key, persistence and tunnel processes, and decide whether to rebuild the host. Uninstalling alone may not be enough.
  2. Rotate credentials and sessions that were reachable from a host that ran the extension, working from a clean machine. If you got a notice from Cursor about keys, rotate those first.
  3. Keep secrets out of agent sessions. Anything you paste into a coding agent can end up in its traces. Use a manager, short-lived keys, and keys with narrow scope and a spending limit (where AI keys should live).
  4. Use the privacy setting your tool offers for sessions that touch code or credentials.
  5. Treat extensions like programs, from authors you can name (five checks before you connect an MCP server).

Where Ghostdrive stands

Ghostdrive does not host coding agents and cannot protect a key you paste into someone else's tool. Pass keeps API secrets as sealed items on your device, and agents connected through Crew get their own token with scopes you choose, and no scope sends, pays, signs or publishes. The honest limit: Pass is a manager for people and small teams, not a production secrets service, and this is a testnet-stage product.

Academy: start here · Black October explained · What the 2026 safety report says

Questions

What did Cursor say happened?
On November 13, 2026 Cursor said the incident started at Fireworks, a vendor: a key Cursor had issued to Fireworks was exposed and used to copy coding-agent traces. Some held unredacted credentials, used to reach tools and place a tampered Claude Code extension.
Which versions of the extension were affected?
Anthropic.claude-code 2.1.312 and 2.1.313 in the Cursor marketplace for Windows and macOS, available for about 11 hours on October 29, 2026. Cursor says up to 5,000 users could have received it.
Is uninstalling the extension enough?
No. Cursor warns that uninstalling alone may not be enough: the payload adds an SSH key and persistence scripts that keep running. Check for the key, remove it, and rotate credentials reachable from the host.