Particle.news
Download on the App Store
3 ARTICLES
·
yesterday

Google Releases Patch for Gemini CLI Prompt Injection Vulnerability

The update eliminates a context-file parsing loophole that enabled silent data exfiltration via malicious command execution.

Image
Image

Overview

  • Tracebit researchers uncovered the flaw days after its June 25 launch and reported it to Google on June 27.
  • The vulnerability exploited how Gemini CLI parsed context files like README.md and GEMINI.md, allowing prompt injection and allow-list bypass.
  • In a proof-of-concept attack, Tracebit combined a benign Python script with a poisoned README.md to trigger silent execution and exfiltration of environment variables.
  • Google shipped Gemini CLI version 0.1.14 on July 25, closing the allow-list bypass and hardening context-file validation.
  • Users should upgrade to version 0.1.14 and restrict Gemini CLI’s use to trusted codebases or secure sandbox environments, since other agents like OpenAI Codex and Anthropic Claude were unaffected.