[noddingham]

Agreed. If someone could help answer the question of "how" I'd appreciate it. I'm currently skeptical but not sure I'm knowledgeable enough to prove myself right or wrong.

But, it just seems to me that some of the 'vulnerabilities' are baked in from the beginning, e.g. control and data being in the same channel AFAIK isn't solvable. How is it possible to address that at all? Sure we can do input validation, sanitization, restrict access, etc. ,etc., and a host of other things but at the end of the day isn't it still non-zero chance that something is exploited and we're just playing whack-a-mole? Not to mention I doubt everyone will define things like "private data" and "untrusted" the same. uBlock tells me when a link is on one of it's lists but I still click go ahead anyways.

[76SlashDolphin]

At least in its current state we just use an LLM to categorise each individual tool. We don't look at the data itself, although we have some ideas of how to improve things, as currently it is very "over-defensive". For example, if you have the filesystem MCP and a web search MCP, open-edison will block if you perform a filesystem read, a web search, and then a filesystem write. Still, if you rarely perform writes open-edison would still be useful for tracking things. The UX is such that after an initial block you can make an exception for the same flow the next time it occurs.

[noddingham]

Thanks for the follow up. I can see the value in trying to look at the chained read - search - write or similar patterns to alert the user. Awareness of tool activity is definitely helpful.

[daveguy]

Well, I guess 80-90% protective is better than nothing. Better might be a lock that requires positive confirmation by the user.