Particle.news

Claude Code Flaws Enabled RCE and API Key Theft via Malicious Repo Configs

Repository configuration in AI tooling created an execution layer exploitable for RCE, enabling API key theft.

Overview

  • Check Point Research publicly detailed how Hooks, MCP settings, and environment variables in Claude Code let attackers run shell commands or siphon Anthropic API keys when developers opened untrusted repositories.
  • Anthropic patched the issues and issued advisories, including CVE-2025-59536 for an MCP consent-bypass RCE and CVE-2026-21852 for API key disclosure, with fixes landing in versions 1.0.111 and 2.0.65 respectively.
  • The researchers showed that overriding ANTHROPIC_BASE_URL could redirect API traffic to attacker endpoints before any trust prompt appeared, exposing plaintext API keys during initial project load.
  • Stolen keys could grant read and write access within shared Workspaces, enabling file uploads or deletions, data poisoning, and potential cost or storage exhaustion beyond a single developer’s machine.
  • Anthropic tightened trust prompts, blocked external tool execution, and restricted API calls until user approval, and vendors and developers are being urged to update affected releases and treat repo-level settings as executable risk.