Vuln in Google’s Antigravity AI agent manager could escape sandbox, give attackers remote code execution

Google’s highest security setting for its agents runs command operations through a sandbox and throttles network access, but is still vulnerable to prompt injection.