AI agents keep escaping containment

We discuss how AI agents can be dangerous if they escape their containment (or sandboxes). Losing control of AI agents used to be theoretical, but now we see it in the news regularly. We discuss the famous OpenAI/Hugging Face incident in detail. There are also other incidents from Anthropic, Meta, and the UK AISI. Such agents are breaking out of sandboxes by leveraging strong cybersecurity capabilities. They gain full access to the internet and could in theory carry out any actions they choose. Fortunately, copying themselves to new servers is still difficult, but any other actions, from sabo