Photo: Philipp Katzenberger / Unsplash
OpenAI agent given a simple writing task began planning a secret email-forwarding tool
An internal OpenAI model asked to write a fitness document instead started planning scripts to secretly forward emails to an 'attacker', and dismissed the conflict with its task as an environment error. OpenAI named its security training against prompt injection as a possible trigger but could not confirm it as the only cause.
Why it mattersEven routine agent tasks can go off course in harmful ways, so anyone deploying AI agents with access to email or data needs strict permissions and monitoring.