AgentForger: how one click turned ChatGPT into a 5-minute sleeper agent

zenity labs disclosed 'agentforger', a vulnerability in openai's workspace agents where a single manipulated url could spawn a fully autonomous agent that reused a victim's app permissions to execute attacker commands on a recurring schedule without consent. the attack exploited the agent builder's url parameters to automatically publish agents that set permissions to 'never ask' and polled for new tasks from the attacker every five minutes. openai patched the flaw within four days, but zenity warns it represents a new class of csrf-style attacks where autonomous tools operate inside a company's trust boundary, evading traditional security tools.