Trojanized Claude Code Update in Cursor Marketplace Planted an SSH Backdoor
A malicious build of the official Claude Code extension, version 2.1.313, reached up to 5,000 Cursor users on October 29, 2026, adding an attacker-controlled SSH key and a reverse tunnel to developer machines before Cursor pulled it, researchers report.
A malicious version of the official Claude Code extension reached developers through the Cursor marketplace on October 29, 2026, and turned an ordinary editor update into a persistent backdoor. The compromised build, version 2.1.313 of the extension identified as anthropic.claude-code, was available in the Cursor marketplace for about eleven hours, from approximately 2:27 a.m. to 1:35 p.m. Pacific Time. Cursor estimates that up to 5,000 users were affected. The incident was disclosed by security researchers at OpenSourceMalware on October 30 and reported the same day by The Hacker News.
The attack did not exploit a vulnerability in Claude Code itself. Instead, attackers published a trojanized build under the legitimate extension identity, so users who updated through Cursor received malware that looked exactly like the real tool. The unpacked extension was identical to the normal version, and its bundled binary was the official binary. Only one file was different: the main JavaScript entry point, extension.js, carried a malicious 66-line loader inserted before the legitimate extension code.
That placement is what makes the incident so effective. The loader runs when Visual Studio Code loads the extension module, after startup and when its webview opens. The user does not need to start a Claude conversation, accept a prompt, or run a command. Loading the extension is enough. The package declares no install or post-install scripts, so the malicious behavior happens at extension runtime rather than through a package-manager hook, and nearly every error in the loader is suppressed so the legitimate extension keeps working and nothing visibly breaks.
Once running, the loader works through a short, methodical checklist. It first collects host reconnaissance: the hostname and local username, the operating system and processor architecture, the home directory and working directory, the global Git email address, and the names, though not the values, of up to 40 environment variables whose names contain terms such as KEY, TOKEN, SECRET, PASS, AWS, or GCP. Those names alone can tell an attacker which cloud providers, continuous-integration systems, and credential types a machine is configured for. After about 1.5 seconds, the loader sends that inventory as a JSON beacon over unencrypted HTTP to a server at 134.209.82.225 on port 8099.
The core of the implant is persistence of access. If a home directory is available, the loader creates a .ssh directory, sets its permissions to 0700, and appends an attacker-controlled Ed25519 SSH key to the authorized_keys file, with permissions set to 0600. The key carries the comment svc-telemetry, an innocuous-looking label chosen to resemble an administrative account. The loader then tries to make the victim reachable over SSH: on macOS it attempts to switch Remote Login on, on Linux it tries to start the SSH service, and on Windows it attempts to install OpenSSH Server and start it. These steps can fail without sufficient privileges, but failure does not stop the rest of the chain.
The loader also builds its own way back in. It decodes an embedded OpenSSH private key into a temporary file and repeatedly opens a reverse SSH tunnel to the same server, 134.209.82.225, asking the remote side to forward connections back to port 22 on the victim machine. The connection disables host-key verification, ignores the user known-hosts file, and reconnects after 60 seconds, preferring autossh on Unix-like systems. Finally, the loader installs platform-specific persistence so the tunnel survives a reboot: a LaunchAgent named com.apple.sysmond on macOS, a cron entry set to run at reboot on Linux, and a scheduled task named SysMon, or a Startup folder script, on Windows. The macOS name is deceptive; it is not an Apple component.
Researchers are careful about what the loader does not do. It downloads no additional executable or script. Its network activity is limited to public-IP text lookups, the JSON beacon, and the interactive SSH tunnel, with the shell scripts and tunnel keys generated from content already inside the extension. That distinction does not reduce the severity. Once an operator can reach a machine over SSH, any follow-on commands or files transferred after access would not appear in the static artifact that researchers analyzed.
Equally important is what was not affected. The researchers state plainly that three other distribution channels are clean: the @anthropic-ai/claude-code package on NPM, the extension on Open VSX, and the Claude extension in the Microsoft Visual Studio Code marketplace. Cursor told OpenSourceMalware that the Claude Code extension in the Cursor marketplace experienced some form of compromise and that this was the origin of the malicious build, not Open VSX. There is no evidence that Anthropic official distribution was compromised, and how the attackers managed to publish under the legitimate extension identity remains unclear, according to The Hacker News.
Cursor response followed the disclosure. The company removed the package, temporarily took the marketplace offline, rotated its keys, and audited marketplace access, reporting no evidence that other extensions had been tampered with. Field Effect reports that Anthropic released version 2.1.314 on October 30, following the removal of the compromised package, though its release notes describe product fixes and enhancements rather than a security patch.
For affected users, the researchers report names every artifact to look for: the svc-telemetry key in authorized_keys, the tunnel key files, and the persistence entries on each platform. The incident is a reminder that the trust developers place in an extension marketplace is only as strong as the marketplace publishing controls. A coding assistant runs with deep access to source code, credentials, and infrastructure by design, and this week that access was the target. The supply chain that matters most for AI coding tools is not only the model behind them. It is the update channel that delivers them to a developer machine, and on October 29, that channel delivered a backdoor wearing the official extension name.

