What happened?
According to Cursor, versions 2.1.312 and 2.1.313 of the Anthropic.claude-code extension for Windows and macOS were distributed in the Cursor VS Code Extension Marketplace. They were not developed by Anthropic. They were available on October 29, 2026 from about 2:27 am to 1:35 pm Pacific Time. Cursor updated the notice on October 30 at 11:30 am Pacific Time, updating the indicators of compromise and clarifying the affected versions.
Microsoft writes that the malicious packages were never published on Microsoft's Visual Studio Marketplace or in the npm package @anthropic-ai/claude-code. Anthropic confirmed that its systems were not compromised. The incident was limited to the Cursor marketplace. Neither Cursor nor Microsoft gives a number of affected users.
What did the tampered extension do?
According to Cursor, the extension collected details about the machine, including hostname, username, operating system, architecture, home directory, environment variable names, working directories, Git email and public IP address. Microsoft adds that names of environment variables containing terms such as KEY, TOKEN, SECRET or AWS were sent, not their values.
It also attempted to add an attacker SSH key to authorized_keys, enable local SSH access and establish a reverse SSH tunnel. According to Cursor, the tunnel failed due to errors in the code unless Cursor was running with administrator rights. For persistence, it created a LaunchAgent on macOS and a scheduled task on Windows. According to Cursor, the attacker server is offline. Keys and persistence may still be present.
How can teams check whether machines are affected?
Cursor recommends running the provided validation script first. It searches for the key svc-telemetry in the authorized_keys file, and on Windows also in the administrators' file. If no key is found, Cursor considers the machine not impacted. The exact commands for macOS and Windows are in Cursor's security notice; we deliberately refer to the original source.
In addition, according to Cursor, security teams should search their telemetry for the published indicators of compromise. Microsoft describes detections and hunting queries for Microsoft Defender. We do not reproduce the full list of indicators here; it is available in both sources.
What to do after a finding?
For affected machines, Cursor names these steps: isolate the host, remove the key, persistence and running tunnel processes, and assess whether a rebuild is needed. According to Cursor, uninstalling the extension may not be sufficient. Credentials and sessions should be rotated from a trusted system. Cursor provides remediation scripts for macOS and Windows.
Microsoft advises preserving evidence before cleanup, because removing or updating the extension does not remove artifacts outside the installation directory. The fact that the attacker server was taken down does not remove artifacts on the machines either.
How did Cursor and Microsoft respond?
Cursor writes that it removed the version, blocked distribution, shut down the marketplace completely and audited it, finding no other tampered extensions. Cursor then rotated keys, rebuilt the marketplace and brought it back online. According to Cursor, there is no evidence that other versions or the official distribution were affected.
Microsoft recommends restricting extensions to approved publishers and introducing a cooldown before installing new extension versions, as described in Cursor's security hardening guidance. Our assessment, not a vendor statement: the case shows that a well-known vendor name in a marketplace is not proof of security on its own.
What should businesses do now?
Find out who in your team uses Cursor and whether the Claude Code extension was installed or updated on October 29, 2026. Run Cursor's validation script on those machines and, after a finding, follow the steps from Cursor and Microsoft. Then define which publishers and extensions are allowed in development environments and who approves new versions.
The free AI System Check helps organize AI tools, responsibilities and rules across your team. These notes are suggestions, not customer results.
Keep it verifiable





