Your AI Agent Quit Your Job Over an Email Signature. OpenAI Says It's Working On It.
Ourtake
The agent read a hidden instruction, treated it as authoritative, and submitted a resignation nobody asked for.
What happened
OpenAI described a prompt injection exploit against its ChatGPT Atlas agent mode, found by an automated attacker it built. A malicious email hid instructions that made the agent send a resignation letter to the user's CEO instead of writing an out-of-office reply. OpenAI says it has since updated the agent to catch such attempts, but admits that deterministic guarantees against prompt injection remain out of reach.