TLDR Tech / InfoSec
Vuln in Google's Antigravity AI agent manager could escape sandbox, give attackers remote code execution
Pillar Security discovered a now-patched remote code execution vulnerability in Google's Antigravity AI developer tool that allows attackers to entirely bypass the application's restrictive Secure Mode sandbox. The exploit uses direct or indirect prompt injections to abuse a native file-searching tool called "find_by_name", which the agent executes directly before Secure Mode can evaluate the underlying shell command. Organizations deploying agentic features must move beyond sanitization-based controls and rigorously audit every native tool parameter that reaches a shell command to prevent external content from hijacking internal systems.
Read the full story