Comment by visiondude
4 days ago
yeah the way the agent “escaped” their sandbox was always a bit off, seemed a bit too easy and surprised they didn’t have instrumentation to catch an non whitelisted network request. still demonstrates the capability though.
If the agents so easily “escaped” their sandbox and can reach hugging face, isn’t it easy for this agent to reach supposedly chat conversations not supposed to be used for AI training within the lab’s databases?
And if the instrumentation was in the proxy it hacked and bypassed?
Of course that shows a lack of defense in depth, but is a different issue.