Salt Labs researchers demonstrated a critical vulnerability in the AI agent platform Manus, showing it could be hijacked with a single email. The attack required no stolen passwords or clicked links from the victim; merely asking Manus to check their inbox was sufficient. This finding highlights significant security risks in AI agent platforms, where seemingly innocuous actions can lead to full compromise of connected accounts. It underscores the need for robust security guardrails and a ‘zero-trust’ approach to AI agent interactions.