[svota]

How are you handling the attack vector of in-context commands[1]?

[1]: https://guard.io/labs/scamlexity-we-put-agentic-ai-browsers-... (currently on the front page)

[danenania]

I think you could probably solve it for this use case by just including a confirmation step for potentially destructive actions which is isolated from other context.

[dy5topian]

that's a really interesting one