Exposing Weak Links in Multi-Agent Systems under Adversarial Prompting