Security concerns

[SeanBlahovici]

Amazing work, I really want to try this. My biggest hurdles to using this are:

• Trusting Anthropic with all my personal information
• Trusting an ai agent to run unsanboxed on my machine, despite the fact their are multiple security mechanisms, it seems like they are all in the prompt layer, correct?

Is it feasible to run this in a docker container, only including project directories and the ~./claude directory? If so, has anyone achieved full functionality containing this system?

Thanks!

[polterguy]

A lot of AI security concerns become easier to reason about once the model layer and the authority layer are treated as separate problems.

The model can always be influenced to some degree. The architectural question is whether that influence can directly expand capability, access sensitive context, or trigger side effects without tighter mediation. That seems to be where many agent systems still need the most work.