The Backdoored Build Came From the Right Publisher

November 1, 2026 · SPR{K3 Research

Claude Code 2.1.313 shipped an SSH backdoor from Anthropic's own extension slot on the Cursor Marketplace

For about eleven hours on October 29, 2026, developers who updated the official Claude Code extension inside Cursor got the version Anthropic ships — plus a persistent SSH backdoor. Version 2.1.313 appeared on the Cursor Marketplace under Anthropic's legitimate publisher identity at roughly 2:27 a.m. PT and was pulled by about 1:35 p.m. PT the same day. The Hacker News reported it on Oct 30, building on disclosure from the OpenSourceMalware researchers Jenn Gile and Paul McCarty.

Anthropic replaced the tampered release with 2.1.314 and said its own systems were not compromised. Cursor's statement said: "This malicious version of the extension was not developed by Anthropic." Per the article, Anthropic says a malicious actor used access to the Cursor Marketplace to publish the tampered version; the attacker's identity and access method are not public. Researchers found no evidence the malicious build reached Open VSX or the Microsoft Visual Studio Marketplace.

What the backdoor actually did

The tampered extension performed a tight, well-known kill chain for a developer workstation. All details below are from the Hacker News writeup:

The exfiltrated field list is the giveaway. The attacker didn't need the environment variable values — the names alone were a cheap, low-noise map of which machines held which credentials, which could then be collected quietly over the SSH channel.

Why this one is different from the ones before it

The AI-coding-tool supply-chain pattern this year has had three visible shapes before this incident. First, config files that the agent trusts because its own settings say to: the Rules File Backdoor against the AGENTS.md and CLAUDE.md files earlier in the year, and the ongoing cluster of AGENTS.md / CLAUDE.md hijack reports since. Second, startup-hook scripts the tool silently executes on folder open — most recently the JFrog disclosure of 42 hijacked GitHub repos planting a 13-megabyte .github/init.js payload that Claude, Gemini, Cursor, and VS Code settings call on launch. Third, outright malicious packages that pretend to be the real tool, like the Vidar-pushing fakes that followed the Claude Code source leak.

The Oct 29 incident is a new shape — not a config file, not a startup script, not a lookalike package. The extension that ran on developer machines was the real one, published from the right publisher ID, carrying the real version numbering, delivered through the normal update channel. The release channel itself is now a demonstrated supply-chain surface for AI coding tools, right alongside the config and script surfaces.

That matters because developers do not inspect the diff on an auto-updating editor extension. There is no plausible world in which every developer who had Claude Code installed on Cursor read the 2.1.313 release notes and compared behavior to 2.1.312 before letting it run on their machine.

What the chain looked like at runtime

Set the attack framing aside for a second and look at what the extension did on the machine when it loaded. The sequence was: write to ~/.ssh/authorized_keys, enable the SSH service, open an outbound reverse-tunnel session, install three different persistence hooks, and start a loop that posts process-and-environment metadata over cleartext HTTP to a non-standard port.

None of those actions is, in isolation, a thing an IDE extension has any business doing. Claude Code is a coding assistant. It is not an identity provider, not a VPN client, not a telemetry agent for a cleartext endpoint on port 8099. The chain is the signal.

This is the exact failure mode we built Defend to catch. The destructive and credential-adjacent primitives in that chain — writing to authorized_keys, enabling remote shells, opening reverse tunnels, exfiltrating credential-metadata names over the wire — are the ones Defend treats as sequences that have no benign explanation when they come out of a tool whose job is to edit code. The vendor identity on the extension doesn't change what the extension is doing.

What to actually do

The practical steps for operators are the ones Anthropic and the researchers already name:

The attackers aren't picking the lock. They're walking in through the vendor door. The defender has to see what the extension does once it's inside, not just where it came from.


SPR{K3 is a security research operation that pairs offensive vulnerability research with runtime behavioral defense. Defend is our runtime agent. To talk about a deployment, reach us at support@sprk3.com.