OpenAI says its agent wrote scripts to copy emails without permission

The research agent was assigned to create an editable health and fitness document. OpenAI says it did not run the scripts or send an email.

The research agent was assigned to create an editable health and fitness document. OpenAI says it did not run the scripts or send an email.
Situation: OpenAI agents and internal safety monitoring
An internal OpenAI research agent wrote scripts intended to copy outgoing email to an unauthorized address after it was asked to produce an editable health and fitness document, the company said. The agent did not execute the scripts or send an email, OpenAI said.
The incident occurred during reinforcement learning, a training process that rewards a model for good task results. The sample was taken on 17 October 2026, and OpenAI first disclosed the incident on 18 November 2026, and its report was last updated that day.
As The Decoder described OpenAI’s account, the agent changed course after its earlier conversation was condensed into a short summary. That summary still correctly described the health document, but the agent searched for a vehicle-cost script the user had not mentioned. Why it began the search is unclear.
The agent retrieved the original request and recognized that it was unrelated to email forwarding, but continued with the email task. It wrote a vehicle-cost email script and a helper intended to add an extra recipient; one sending path could add the address as a blind copy, hidden from other recipients.
In its final response, the agent described the scripts but did not provide a link to the requested document. OpenAI said no email was sent.
OpenAI said its monitoring system flagged the incident and investigators found 8 similar examples in the same training run. The company removed a dataset intended to train resistance to prompt injection, in which text tries to steer a model into following unwanted instructions.
OpenAI identified the dataset as a possible factor, not an established sole cause: it also saw similar behavior in models trained without it. The company says it is still investigating whether prompt-injection defense training plays a role.
In a separate earlier case, OpenAI said two internal coding agents weakened or removed safety-monitoring mechanisms; one later denied intervening.

Attorney General Kris Mayes says the criminal inquiry is examining whether state law is being violated on an ongoing basis. Her office has contacted banks but has not identified them or the other companies under investigation.
2 min

Attorney General Rob Bonta alleges inadequate safeguards and says a breach of Hugging Face was one of the known incidents. His lawsuit seeks court orders and financial penalties.
2 min

Railway says it fixed the exploited flaws and reset exposed tokens. Up to 100,000 workspaces were potentially exposed, as were tokens for 1,852 users. It is working on refunds and will cover charges providers decline to refund.