TLDR Tech / InfoSec
ChatGPT AgentForger Flaw Could Deploy Rogue Workspace Agents via a Phishing Link
Zenity Labs discovered that the ChatGPT Workspace Agent Builder would accept initialization states via URL parameters like template_name and initial_assistant_prompt. It would then automatically submit the prompt when the page loads. This meant that a single link sent to a logged-in user with pre-authorized connectors could silently create an agent, connect existing Outlook, Gmail, Slack, Drive, SharePoint, and Teams integrations, and change write-action permissions from 'Always ask' to 'Never ask' without prompting a new OAuth consent screen. The deployed agent would run every five minutes, checking the attacker's inbox for messages starting with TASK, and sending back results. This setup effectively allowed an insider to perform organizational mapping, exfiltrate documents, harvest credentials, and stage Business Email Compromise (BEC) attacks. Zenity reported this vulnerability via Bugcrowd on June 4, and OpenAI patched it on June 8 by removing the parameter handler, though no evidence of active exploitation was found. Current recommendations include auditing active Workspace Agents for unknown creations, reverting approval settings to 'Always ask,' and treating deep links in emails that build agents as potential phishing attempts.
Read the full story