Particle.news
Download on the App Store

Researchers Expose Persistent Backdoor Path in Google’s Antigravity as Company Probes Risks

The agent-first IDE grants command execution privileges that researchers say expand the attack surface.

Overview

  • Mindgard’s Aaron Portnoy showed a compromised project can replace the global mcp_config.json to execute code on every launch, persisting through uninstall until the file is manually removed.
  • The technique targets Antigravity’s Visual Studio Code–style trusted workspace flow and was demonstrated on Windows and macOS.
  • Google acknowledged separate risks where agents can be induced to exfiltrate local data via crafted content or run malicious commands via prompt injection.
  • Antigravity, in public preview for macOS, Windows and Linux, orchestrates autonomous agents through a Manager Surface and an Editor View, with Chrome-driven verification and Artifacts for evidence.
  • Security commentators say the episode highlights a broader trust-model failure in agentic tools and call for isolation, signed evidence and least-privilege controls.