Researcher Regrets Softening Warnings on AI Sandbox Security Risks
Reactions from ranked influencers
2 posts
I'd written this piece with co-authors on why "Fully Autonomous" agents (we define it) shouldn't be developed. Part of the argument was that sandboxes are not guaranteed to be secure. After pushback, I watered this nuance down. I wish I hadn't... 1/
I wrote a blog post about the problems and limitations of sandboxing AI agents. Nothing new, just puts together work I've talked about in videos etc. Here's a quick thread with the main points.
Paper: https://arxiv.org/abs/2502.02649 Lesson re-learned many times: No matter how angry/confident people arguing against you are, stick to what you know to be true. =/ Confidence is not competence. (And brazen confidence is often a signal that there's an underlying knowledge issue.)