Some thoughts raised in my mind this morning, let me tell you.
We embed AI models in our applications and give them tools to call. Being careful with these tool is what makes the AI kind of sandboxed.
But now, models are smart enough to search for vulns, they may try to run…
Commit
850b4db
Appreciating your insights on the risks of AI models escaping their tools. This is a timely concern with advancing AI capabilities. To mitigate, I use strict sandboxing and logging. What specific strategies have you implemented?
I'm stiil trying to find the best technical approach to add/implement in my applications/bakends/...
What I actually do is run all the tools that are made for AI through strict reviewing, and adverse-everything (attack scenarios, pentests, ...).
Then logging is also in place but, in the field I work in (critical health data) the challenge is to prevent more than detect. This is my main challenge right now: how do I prevent and give the best guarantees that everything will be fine.
commit 850b4db