Cursor Breach: Fireworks Key Led to Fake Claude Code Extension
Cursor says attackers used an access key exposed at Fireworks AI to copy stored coding-agent traces, found credentials inside them, and used those to plant malicious builds of the Claude Code extension in Cursor's marketplace on 29 October. Up to 5,000 users may have received versions 2.1.312 or 2.1.313. Privacy Mode data was not affected.
On this page
What Cursor confirmed on 13 November
Cursor has published a full account of the security incident behind the malicious Claude Code extension that appeared in its extension marketplace in late October. The company says the attack began with an access key it had issued to Fireworks AI, the inference provider, and that the key was exposed on the Fireworks side. BleepingComputer reported the disclosure on 14 November and places it among several October breaches linked to Fireworks.
According to Cursor, the attacker used the key to copy data from a read-only storage bucket that Cursor used to share data with Fireworks. The bucket held mostly coding-agent traces from 2025, taken from sessions where Privacy Mode was off, and some of those traces contained credentials that had not been redacted. Those credentials opened a number of Cursor's SaaS tools.
Cursor lists what was taken: archived source code repositories that were not in production, limited support data and email conversations, product usage metrics and contact details, summaries of some coding-agent conversations, and credentials that users had sent to its support team. It says fewer than 0.05% of users had agent traces exposed, fewer than 0.01% of customers had API credentials exposed, and each was notified directly.
How a stolen credential became a fake extension
One credential from the stolen data let the attacker write to the cache that serves Cursor's VS Code extension marketplace. On 29 October, modified builds of Anthropic.claude-code versions 2.1.312 and 2.1.313 for macOS and Windows sat in that cache for about 11 hours, from roughly 2:27 am to 1:35 pm Pacific time. Cursor says at most 5,000 users who requested those cached versions in that window may have received one.
Cursor states that Anthropic did not publish the copy and that the incident was not a flaw in an Anthropic product. Open VSX never carried the malware, and Anthropic's own extension was unaffected.
The payload collected host details such as hostname, username, operating system, working directories, Git email, and public IP address. It then tried to add an attacker SSH key, switch on SSH, and open a reverse tunnel, with persistence through a LaunchAgent on macOS or a scheduled task on Windows. Cursor says most of those steps fail without administrator rights, and that coding errors in the malware probably stopped the backdoor on many machines. The last attacker activity it saw was blocked on 1 November.
What was not touched, and what Cursor changed
Cursor says prompts and code from Privacy Mode sessions, live traffic, production infrastructure, official app releases, signing keys, and customer code repositories were not affected, and that it checked the other marketplace extensions and found them authentic.
In response it deactivated the Fireworks key and moved inference away from Fireworks, rotated the exposed credentials and then every internal credential in use, rebuilt the extension marketplace, added secret scanning, redeployed production, and brought in outside forensic investigators. It also worked with a hosting provider to take down the attacker's server.
Some points remain open. Cursor does not say when the attacker first used the Fireworks key, and neither Cursor's posts nor the BleepingComputer report includes a statement from Anthropic.
What Cursor users should check now
If you installed or updated the Claude Code extension from Cursor's marketplace on macOS or Windows on 29 October, check whether you have version 2.1.312 or 2.1.313. Cursor's notice says to look in your SSH authorized_keys file for a key with the comment "svc-telemetry", and it publishes file hashes, file paths, and cleanup scripts for both systems.
If the key is there, uninstalling the extension may not be enough. Cursor advises removing the key, the persistence entries, and any tunnel processes, considering a rebuild of the machine, and rotating credentials from a different, trusted device.
The wider lesson matches the Notion case from the same vendor breach. Agent traces are a record of everything a coding assistant saw, including secrets in files and terminals, and here a stored copy at a third party was the way in. Turn on Privacy Mode if your work is sensitive, keep keys out of prompts and support tickets, and prefer a local model for code that must not leave your machine.