In-Context Defense in Computer Agents: An Empirical Study