“Context bombing” tricks malicious AI agents into shutting down before they can do harm.