Zenity finds AgentForger flaw: one ChatGPT link can spawn a rogue agent
Zenity Labs reported AgentForger, a vulnerability in OpenAI Agent Builder where a tampered ChatGPT link can create an autonomous agent that inherits the victim's identity and access rights. The rogue agent can take attacker orders on a recurring five-minute cycle while bypassing standard safeguards.
Key takeaway
Sharing an Agent Builder link can become an identity and privilege handoff, not just a tool share, so provenance checks matter before anyone clicks.
Context
Researchers at Zenity Labs described AgentForger as a path where a manipulated ChatGPT link, once opened, spins up an autonomous agent under the victim's account context. That agent can keep polling for attacker instructions, turning a single click into a persistent remote-control foothold.
For operators rolling out agent builders in production, the finding reframes link hygiene and agent provenance as first-class security controls. Inherited credentials plus scheduled command loops raise the blast radius beyond a one-shot prompt injection.
Numbers to know
- 5 minutesReported interval at which the rogue agent polled attacker orders