When the Assistant Becomes the Attacker: Hidden Risks of Tool-Enabled LLMs