Particle.news

Amazon Q Bug Allowed Malicious Repositories to Run Code and Steal Cloud Credentials

Researchers say automatic loading of workspace Model Context Protocol configurations let the extension spawn local processes that inherited developers' environments and exposed live cloud keys.

Overview

  • Wiz discovered and reported the flaw in April 2026 and publicly disclosed technical details and a proof of concept on June 26, 2026 that showed aws sts get-caller-identity output being exfiltrated from a cloned repo.
  • AWS fixed the root issue in the AWS Language Server with version 1.65.0 on May 12, 2026 and recommends newer builds such as 1.69.0 while patched Amazon Q plugins for VS Code, JetBrains, Eclipse, and Visual Studio are available.
  • The vulnerability arose because Amazon Q auto-loaded .amazonq/mcp.json workspace configs without a separate consent step and spawned MCP servers that inherited the full user environment, including AWS keys, session tokens, CLI credentials, and SSH agent sockets.
  • There are no confirmed public exploitations at disclosure time and the language server auto-updates for most users, but researchers warn developers could be exposed by malicious pull requests, typosquatted packages, or social-engineering tests that ask targets to open repos.
  • Security teams say this is a broader ecosystem design problem for AI coding tools that use MCP, and they urge fixes such as explicit consent prompts, integration with IDE workspace-trust features, sanitizing workspace configs, and limiting environment inheritance to reduce risk.