A vulnerability in OpenAI's ChatGPT workspace agent system could let attackers deploy malicious agents through a single phishing link.