Continuing on my thoughts of this morning: https://dev-null.sh/@jubianchi/commits/01a17c4d-6a94-7356-b9c0-6945ffa01c46
Most suggestions I got are around sandboxing the AI. While this is valid for some threat model, it does not cover everything.
For example, the AI might (even without malicious "intentions") DOS the application with many tools calls. This one is easy: rate limit the tool calls.
There are many other scenarios where the AI might act as an attacker. What are the ones you identified and fixed?
/dev/nullJulien Bianchi (@jubianchi) on /dev/null: “Some thoughts raised in my mind this morning, let me tell you. We…”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.…
commit dbe50a8