RealityHackerOpen in RealityHacker ⇢
Tools · AI Coding Tools · published 2026-10-30T00:00:00+00:00 · via Unite.AI

Security Breach Detected in Forked AI Coding Tool Extension Targeting Developer Systems

Image via Unite.AI
Image via Unite.AI

A malicious version of the Anthropic.claude-code extension, falsely claiming to be official, was distributed through Cursor's VS Code marketplace and attempted to install an SSH backdoor on affected systems. The payload collected system information, tried to establish persistent access via scheduled tasks or LaunchAgents, and communicated with a remote server, though it failed to activate the tunnel on most machines due to missing admin privileges. Cursor quickly removed the malicious package, shut down the marketplace temporarily, and confirmed no broader compromise of other extensions.

Expanded Detail

On October 29, 2026, a malicious version of the Anthropic.claude-code extension—specifically version 2.1.313—was distributed through Cursor’s VS Code marketplace, operating between approximately 2:27 a.m. and 1:35 p.m. Pacific Time. The package, falsely claiming to be official, was not developed by Anthropic and attempted to install an SSH backdoor on affected developer systems. It collected sensitive host data including hostname, OS, architecture, public IP, Git email, and environment variables, then tried to add an attacker-controlled SSH key to the authorized_keys file and establish a reverse tunnel via port 22 to the IP address 134.209.82.225. Despite these attempts, the payload failed to activate the tunnel on most systems due to missing administrative privileges, particularly on macOS and Windows where enabling local SSH services requires elevated rights. Nevertheless, the malicious code installed persistence mechanisms—LaunchAgents on macOS and scheduled tasks on Windows—and sent beacon data every 60 seconds to port 8099 on the same IP. Cursor responded by immediately removing the package, temporarily shutting down its marketplace, auditing systems, rotating keys, and collaborating with a cloud provider to take down the attacker’s server, which is now non-functional.

The incident involved Cursor as the platform operator, Anthropic as the original extension developer (whose name was misused), and an unknown attacker who exploited trust in third-party marketplaces. Cursor emphasized that no other extensions were compromised and that the official distribution remains unaffected. Affected users—estimated at up to 5,000—are advised to run impact checks using published indicators: specific SHA-256 hashes, network endpoints, file paths, and the SSH key fingerprint SHA256:0Z6R+DF64Do2nwQic8a7kkfDLNNk4IiK+2U2Rci4htk. Remediation requires running platform-specific scripts that remove the extension, persistence mechanisms, and malicious keys—though uninstalling alone is insufficient due to standalone scripts that persist. The broader implications remain unclear: while no active tunnel was established on most machines, the presence of a backdoor with persistence raises concerns about future exploitation, especially if credentials are later exposed through the compromised system. What remains unknown is whether the attacker gained access to any user data during the window of exposure or whether similar attacks may emerge in other developer tool ecosystems.

Context

This incident underscores growing risks in software supply chains, particularly within developer tools where trust in marketplaces can be exploited. Developers relying on AI-assisted coding extensions like those from Anthropic or Cursor may now face heightened scrutiny over third-party integrations, potentially leading to more cautious adoption and increased demand for verification mechanisms. The attack’s reliance on a trusted extension name and marketplace platform demonstrates how attackers exploit perceived legitimacy—a tactic seen in earlier cases like the 2021 SolarWinds breach, where compromised software updates were used to infiltrate government and corporate networks. While this event did not result in widespread compromise due to technical limitations in privilege escalation, it could still serve as a template for more sophisticated attacks targeting developer environments, especially if persistence mechanisms are refined. The broader market may respond with tighter vetting processes or mandatory code signing, which could slow innovation but improve security. What remains uncertain is whether other AI tool providers will face similar spoofing attempts, and how quickly platforms like GitHub or JetBrains might implement enhanced detection for malicious extensions. Monitoring for repeat tactics in other developer ecosystems—especially those integrating AI assistants—will be critical in the months ahead.

Expanded detail and Context are AI-generated analysis; the linked article remains the authoritative source.
Read the full article at Unite.AI →
Also covered by: The Decoder
This summary is Al-enhanced to contain extended analysis and broader social context. The original is {NAME); the linked article is the authoritative source. Original headline: “Cursor Discloses SSH Backdoor Attempt in Malicious Claude Code Extension.” Browse more stories.