Itdaily - Cursor confirms how leak led to malicious Claude Code extension

Cursor confirms how leak led to malicious Claude Code extension

Smartphone showing the Cursor logo in front of a screen with the website of the Cursor AI code editor.
Image: bella1105/Shutterstock

A leaked access key at partner Fireworks gave an attacker the chance to place a malicious Claude Code extension in Cursor’s marketplace. According to Cursor, at most 5,000 users received the infected version. 

Cursor, the AI code editor from Anysphere, has confirmed after two weeks of investigation how an attacker got a tampered copy of Anthropic’s Claude Code extension into its VS Code extension marketplace on 29 October. The affected versions are 2.1.312 and 2.1.313 for Windows and macOS. Anthropic did not publish those versions, and according to Cursor there was no vulnerability in Anthropic’s product. 

The extension was online for about eleven hours, between 10.27 am and 9.35 pm Belgian time. At most 5,000 users who requested exactly those versions during that period received the malicious build. Open VSX, the source that Cursor’s marketplace draws from, never contained the malware. 

Leak started at Fireworks 

The attacker gained access to a key that Cursor had issued to Fireworks, an external partner for training and inference. Fireworks leaked that key. The attacker used it to copy data from a storage bucket that Cursor shared with Fireworks: traces of coding agents, mainly from 2025, from sessions that were not running in Privacy Mode. 

A small portion of those traces contained credentials that were not protected. For less than 0.01 percent of customers, API keys were exposed. With that data, the attacker gained access to a few of Cursor’s SaaS tools, including archived source code that is not in production, support conversations and usage statistics. With a stolen credential from those systems, he placed the fake extension directly in the extension marketplace’s cache. 

What the Claude Code extension did 

The infected extension collected data about the device and the user, such as hostname, username, operating system, public IP address and the configured Git email address, and sent it to the attacker. It also added an SSH key belonging to the attacker to authorized_keys and tried to set up a reverse SSH tunnel, which came back after every restart via a LaunchAgent on macOS or a scheduled task on Windows.

Administrator rights are needed to enable the local SSH service. On most devices, the extension did not have them. According to Cursor, the attacker therefore could not log in through the tunnel there. The key and the background scripts remained on those devices. 

Extension taken offline 

Cursor took the extension offline as soon as it discovered the infection and then shut down the entire marketplace. Together with the attacker’s cloud provider, it took the attacker’s server offline. Cursor checked all other extensions for authenticity and rebuilt the marketplace. 

The company deactivated the Fireworks key and moved all of its inference traffic to other providers. All internal credentials were replaced, even where there was no trace of abuse. Cursor added scans for secrets in code, restricted network access at its service providers and redeployed the entire production infrastructure. External forensic experts helped with the investigation. According to Cursor, the Cursor app, the official updates, the production systems and customers’ code repositories were not affected. 

Cursor considers the incident contained. The attacker’s last activity dates from 1 November and was blocked. Affected customers were notified personally. Anyone who installed version 2.1.312 or 2.1.313 on 29 October needs to do more than remove the extension: Cursor provides scripts that find the key labelled svc-telemetry and clean up the leftover scripts, and recommends replacing credentials on affected devices.