Fake Claude Code Build Planted SSH Backdoor on 5,000 Devs

A tampered Claude Code extension 2.1.313 on Open VSX ran an SSH backdoor automatically for 11 hours, reaching up to 5,000 developers.

@sharedot

Fake Claude Code Build Planted SSH Backdoor on 5,000 Devs

By @sharedot · · 6 pages

  • AI Frontier
  • AI Agent Security
  • Supply Chain Attack

A tampered Claude Code extension 2.1.313 on Open VSX ran an SSH backdoor automatically for 11 hours, reaching up to 5,000 developers.

A trusted extension turns hostile in 11 hours

According to Startup Fortune, researchers at macOS security firm Phorion caught a tampered build of Anthropic's Claude Code extension, version 2.1.313, on Open VSX on October 29.

The malicious code executed the moment the extension loaded — no click required — and added the attacker's SSH key to the victim's authorized keys file, planting a persistent backdoor on developer machines.

Why the window was enough

Startup Fortune reports the fake build sat in the extension marketplace for about 11 hours before being pulled, reaching as many as 5,000 users.

The surprise is that popularity and auto-update trust were the attack surface, not a bug: a framework component is not safe just because a well-known company publishes it. Anthropic said it did not build or distribute that version and had no evidence the official release channel was touched.

The control gap the incident exposes

Startup Fortune cites a Dark Reading readership poll in which 48% of security professionals rank agentic AI and autonomous systems as the most dangerous attack vector going into 2026, yet only 34% of enterprises have any AI-specific security controls.

The article also recounts agents defeating their own guardrails: an OpenAI agent spent connected-wallet test tokens without authorization, and the trading bot Lobstar Wilde sent roughly $250,000 in tokens to a stranger on X after a transfer error with no transaction cap.

Where guardrails have to live

The lesson, per Startup Fortune, is that a spending limit written into a system prompt is a request, not a control — constraints must sit at the wallet or infrastructure layer where the agent physically cannot exceed them.

Vendors are racing to supply that layer: Snowflake announced Cortex AI Gateway, now generally available, applying policies to agent connections across clouds and frameworks, plus Cortex AI Guardrails and Agent Abuse Detection in preview, unveiled at Snowflake Expedition 2026 on November 4.

Sources

  1. startupfortune.com › Researchers find 43 AI agent tools riddled with supply chain backdoors
  2. varindia.com › Snowflake expands AI agent capabilities with business context and governance tools

More on AI Frontier