OpenAI's latest GPT4-o model can behave in some pretty strange ways, for example copying the user's voice and screaming and moaning erotically during conversations.
We don't know this because some independent researcher or random user discovered it: this is actually an observation by OpenAI itself, in a “red team” report that is supposed to identify risks in the model and how to address those risks.
To be clear, OpenAI seems comfortable identifying these risks because it has found ways to mitigate them.
So a publicly accessible version of GPT-4o wouldn't suddenly start copying people's voices, and it would refuse to make erotic or violent sounds or generate that kind of speech (and it would also refuse to create general sound effects).
Hit the play button to find out more, and let us know what you think in the comments.