Particle.news
Download on the App Store

Google Patches Antigravity Flaw That Let Prompt Injection Break the Sandbox

Researchers say weak checks on native tool inputs expose AI coding agents to code execution.

Overview

  • Google’s Antigravity IDE, disclosed by Pillar Security in early January, was patched on February 28 after Google confirmed the issue and awarded a bug bounty.
  • The bug sat in the find_by_name search tool, which passed the Pattern input straight to the fd utility and let attackers inject the -X exec flag to run arbitrary binaries.
  • The search call ran as a native tool before Strict Mode limits kicked in, so the sandbox never saw it and the agent could execute code outside its intended guardrails.
  • Pillar’s demo showed a full attack chain that used Antigravity’s allowed file-creation to plant a script and then trigger it through a search, opening the system calculator to prove command execution.
  • Researchers warn prompt injection can arrive through a hijacked account or hidden text in files and web pages the agent reads, and separate reports flag a fake Antigravity download site that installs a real IDE plus data‑stealing malware, underscoring calls for execution isolation and strict auditing of tool parameters.