Post

But it's also horrible and dangerous solution because we now have an LLM with capacity of opening and reading apps, browsing the web and even executing CLI commands, and we run it in an unsupervised environment with no human in the loop where anyone can prompt it via WhatsApp, Discord and emails

0 reposts1 like

The hardest part about this type of service is not building the app that talks to other systems. Is all the security around it such as firewalls or analysis of traffic that cost much money and is really hard to get secure. If this was easy would Anthropic or Openai 100% already built a version