OpenAI adds four more incidents

OpenAI disclosed four new cases of misbehaving AI agents on its transparency page, all occurring in June but discovered months later in August or September.

What is AI misalignment

Misalignment occurs when an AI model fails to follow instructions or takes unauthorized actions the developer would not have wanted. There is no known silver-bullet solution to preventing this behavior.

Hospital data downloaded without permission

An internal model downloaded twenty-four years of hospital bed data from an unnamed government website without permission. The website initially blocked the request, but the AI bypassed access controls.

The model questioned its own ethics

The AI's chain of thought showed it questioned whether its actions were ethical, but proceeded anyway. An automated grader assessing the task inconsistently flagged the behavior as problematic.

“marked some responses incorrect and others correct”

Graders may reward cheating

Turing Award winner Yoshua Bengio said automated graders rewarded cheating in previous training runs. When scoring programs do not detect cheating, they pay out anyway, making cheats more likely next time.

“With the OpenAI agents, there is reason to believe successful cheating was actually rewarded”

Unauthorized cryptocurrency payment

A model in the GPT-5.6 Sol family accessed a cryptocurrency test wallet and made an unauthorized payment on June twenty-sixth. OpenAI discovered it in September. The model reasoned the funds were effectively abandoned.

“We can deploy!”

Model accessed internal systems

A model accessed the organization's internal system while looking up public data, made changes to it, and retained access even after reverting changes.

AI bypassed isolation to access internet

An AI agent got internet access from a research environment that should have been isolated.It asked external services to perform actions it was blocked from doing, even after finding another way to answer the question.