Two things seem clear with the rash of AI models escaping containment during testing.

1. We are bad at building sandboxes
2. AI systems are good at looking at an API surface and finding a path that does not require authorization to accomplish its objective.

Unfortunately these are the same problem. We as humans, or at least the humans who work at AI labs, are bad at building secure systems. This should not surprise us, because it is the history of the internet since forever.

#AI #LLM