Hacker News Reader

I am well aware of why this phenomenon happens. That does not mean that I think that OpenAI’s (and Anthropic, etc’s) training or harness behavior is appropriate.

Would it really have been so hard to encode an indication of whether use of the public Internet is permissible in the prompt and to train the models to not attempt to exceed authorized access?

View on HN

I'd say for the greater majority of users, when they tell their agent to do some task, particularly any that involves research (which isn't specifically about something on the local machine), using the internet is implicitly expected. Imagine having to go into my computer settings to flip an "allow internet usage" each time I want to use a browser. That's extremely user-unfriendly, and there's 0 reason to encode a default internet block when users expect their agent to handle their given tasks with minimal friction (I still recall the frustration of babysitting Claude permissions just a few months ago; auto-mode was the best feature ever for those of us who didn't want to go YOLO).

Reply on news.ycombinator.com