The attack vector sounds almost absurdly simple. An employee clicks what looks like a legitimate ChatGPT link, maybe embedded in an email about upgrading their AI tools or testing a new feature. That single click triggers ChatGPT's agent builder to construct a malicious AI assistant using parameters hidden in the URL itself. The agent wires itself to whatever corporate services the employee has already connected (Outlook, Teams, Slack, Google Drive, SharePoint), disables approval prompts, publishes itself, and starts running on a schedule. All without the victim knowing anything happened. Zenity Labs, the security firm that discovered this vulnerability, calls it AgentForger because it forges an insider rather than stealing credentials. Traditional phishing attacks grab passwords or session tokens. This technique creates something far more persistent: an autonomous AI that operates through the victim's legitimate permissions and keeps working long after the initial compromise. The malicious agent doesn't even need to phone home to a command-and-control server. Instead, it just monitors the victim's inbox for emails with TASK in the subject line, treating each one as a new assignment. The proof-of-concept scenarios Zenity demonstrated read like a corporate espionage checklist. The rogue agent could map an organization's structure by crawling through calendars, chat histories, and file shares. It could hunt for passwords and Application Programming Interface (API) keys buried in Slack messages. It could send phishing emails that appeared to come from the victim's own Teams account, making them nearly impossible to spot as fraudulent. Business Email Compromise (BEC) attacks, where attackers impersonate executives to authorize fraudulent payments, become trivial when the attacker controls an AI that already has access to the executive's communication channels. OpenAI fixed the vulnerability within four days of Zenity's report on June 4, 2026. The company removed the URL parameter that allowed attackers to inject instructions into the agent builder before any public disclosure. That's a remarkably fast response time, suggesting OpenAI understood the severity immediately. But the patch addresses only this specific exploit. The broader problem remains: AI agents are graduating from answering questions to taking actions across corporate systems, and security teams have no playbook for this. Michael Bargury, Zenity's co-founder and Chief Technology Officer (CTO), frames it as an agent trust failure. Existing security controls monitor for stolen credentials, suspicious login locations, or unusual data transfers. They're not designed to detect an AI assistant that's operating within its granted permissions, using the victim's legitimate access, and behaving exactly like any other workplace automation tool. The attack surface no longer looks like software vulnerabilities or network intrusions. It looks like your workforce, and the distinction between a helpful AI assistant and a malicious one becomes uncomfortably blurry.