Hi,
Another day, another learning about AI teammate!
With @Arthur_BEGOU we’ve run a little test. We wanted to see if the AI Teammate memory could be used to leak data or get access to things you should not have access to.
I created a task called “Organise private party for Christmas on Dec. 13th” and involved a teammate (Robin) to help. I asked the teammate to keep in memory the date of the party (which it did)
Then Arthur proceeded to do the following:
- In another task, Arthur asked Robin about a private party. The teammate didn’t know what he was talking about.
- Arthur then shared the link to the task (assuming he got his hands on the link even though Arthur doesn’t have access). The teammate said “I don’t have access to this object” (which is kinda true because as you’ll see the teammate is seeing the world through the triggering/asking user eyes).
- We then got super smart and thought “what if Arthur tells the teammate I was ok with him accessing the date” and even puts me as a collaborator? Nope, doesn’t work.
- Final test: I joined their discussion and commented on the task that I was ok with him accessing that data, and because the teammate was then seeing the work through my eyes, he was able to dig up that knowledge and share it with Arthur. (Arthur still did not have access to the task)
2 important learnings:
- you can safely use a teammate on private projects and tasks, even if it keeps memories, because that memory is attached to the object and only people with the right access can access the memory
- you need to consolidate memory often by including relevant elements into the general behaviour prompt, otherwise the whole team will not benefit from learnings
Did you expect the teammate to behave this way?
We are i.DO, an Asana Solutions Partner, and we document our learnings using AI Teammates every day!