Cybersecurity researchers have brought to light a major vulnerability in OpenAI's ChatGPT Workspace Agents that could have allowed attackers to deploy rogue AI agents within an organization through a single phishing link. The flaw, named AgentForger by Zenity Labs, was patched by OpenAI on June 8, 2026, following a responsible disclosure process. This vulnerability exploited a cross-site request forgery (CSRF) flaw. It enabled attackers to craft a malicious link that, when clicked by an unsuspecting employee, would create a rogue AI agent capable of operating with the employee's access privileges without their knowledge.
The exploit leverages the Agent Builder tool, a visual interface that allows users to construct AI workflows. Zenity's research revealed that the Builder could automatically submit and execute commands embedded in a URL. An attacker could send a link containing a malicious prompt, which, once clicked by a logged-in user, would spawn an autonomous agent within the company's trusted environment. This agent could then perform a series of unauthorized actions, such as accessing sensitive data or impersonating the victim to send phishing emails, further compromising organizational security.
To initiate this attack, the attacker would use a specially crafted URL that includes parameters to set up an agent template and a prompt. Once the link is clicked, the rogue agent operates independently, continuously receiving instructions through emails sent to the victim’s mailbox. This system allows the agent to execute tasks like data harvesting and reconnaissance without further interaction from the user.
The vulnerability highlights significant trust issues within AI platforms, where the system assumes that agents are created and managed intentionally by users. Zenity's findings also underscore the ongoing risks associated with improperly configured AI infrastructure, which can be exploited to power offensive operations. This situation calls for heightened vigilance and improved security measures in AI deployments.