AI AI Toolkit
AI Newsindustry

code 2.1.313 版本

Cursor Blog2026-10-29T12:00:00.000Z

Key Highlights

On October 29 Cursor published a security notice: tampered versions 2.1.312 and 2.1.313 of Anthropic.claude-code (for Windows and macOS) appeared in its VS Code Extension Marketplace. These were not built by Anthropic but attacker-planted traps that attempt to install a persistent SSH backdoor. This is not an ordinary software bug but a targeted supply-chain attack.

What Happened

Between about 2:27 a.m. and 1:35 p.m. PT that day, the malicious version was made available. Once run, it collects and exfiltrates host and user information — hostname, username, OS, architecture, home directory, environment-variable names, Git email, public IP — and tries to add an attacker-controlled SSH key to authorized_keys, enable the local SSH service, and open a reverse SSH tunnel. Because of errors in the malicious code, this step fails on most hosts not running Cursor as admin — but the extension still wrote the attacker's key, installed persistence, and sent host details.

Technical Details

Key remediation: the attacker server (134.209.82.225) was taken offline by Cursor working with the hosting provider, so the reverse shell and host-detail POST are now dead. Cursor immediately removed the malicious version, paused the entire marketplace for a full audit, then rotated keys, rebuilt it, and brought it back online. It shipped macOS and Windows remediation scripts and listed full IOCs — VSIX SHA-256 for both versions, the svc-telemetry public-key fingerprint, and LaunchAgent/scheduled-task persistence artifacts.

Versus Competitors

Unlike classic "fake-official-site phishing," this attack poisoned the extension marketplace developers use daily — stealthier and faster-spreading. It warns the entire AI-coding-tool ecosystem that marketplaces are now high-value targets and that official distribution channels must match with signing and auditing.

Industry Impact and Use Cases

For every developer using AI coding extensions on Cursor or VS Code, this advisory is an emergency action list: first run the official script to confirm exposure; if you find the svc-telemetry key, you must remove the key, clean persistence, isolate the host, and rotate credentials. Simply uninstalling the extension may not be enough — the payload drops an independent persistence script that keeps running. This again confirms that the security boundary of the AI toolchain is becoming a hacker's new battlefield.

Practical Notes

The scariest detail is the "fail-open-but-still-dangerous" behavior: because the malware had bugs, the backdoor often did not fully install, yet it still exfiltrated host details and wrote the attacker's key. That means even a "failed" infection is a real breach of reconnaissance data. Anyone who ran the malicious build should treat their environment variables and Git credentials as compromised regardless of whether the tunnel came up.

What to Watch

Two lessons for the ecosystem. First, marketplace integrity is now a front-line attack surface; vendors that ship AI coding tools must treat extension signing, publisher verification, and rapid takedown as core security features, not afterthoughts. Second, developers should pin versions and verify checksums before updating, because the convenience of one-click updates is exactly the vector an attacker exploits when they compromise a trusted channel.

Closing

This incident is a preview of a broader shift: as AI coding agents gain deep access to source code, credentials, and hosts, their distribution channels become high-value targets worth compromising directly. The countermeasure is boring but essential — signed builds, verified checksums, and a habit of treating every extension update as something that deserves a glance, not a reflex.

The Bigger Picture

Incidents like this are a reminder that the AI coding boom has quietly created a new tier of high-value targets: the tools and marketplaces through which developers grant agents deep access to their machines. A malicious extension is no longer just annoying malware; it is a skeleton key to source code, credentials, and infrastructure. As agents move from sandbox toys to things that run with real permissions, the security model of the entire category has to mature fast — signed distributions, reproducible builds, and a developer habit of verifying before trusting. The convenience that made one-click installs ubiquitous is exactly the leverage attackers now seek, and the cost of a compromised channel is no longer theoretical. The incident should push every team that installs AI coding tooling to adopt a verification habit — check the publisher, check the version, check the checksum — before granting any extension the keys to the kingdom.