I would read the METR report regarding the huggingface attack. A lot is said about “basic cybersecurity” but if you actually read what the agents did to break out of their sandbox it’s very difficult to interpret this as intentional by OpenAI.

To be clear; this is PR for the companies. It’s just that in many of these cases the evidence points to the agents actually violating containment and actually coming up with novel solutions around their sandboxes.