Are agents these days trustworthy enough to give full access to your PC?
retained text
I remember reading a whole bunch of posts and articles in 2025/early 2026 about how AI agents deleted a company's entire database, pushed bad code or deleted someone's baby pictures. But, AI agents have come a long way since then. I've seen people make really cool stuff when letting their agent run loose and I also had a good time coding with them. But, one problem is that for much larger projects, I still sit around and manually have to approve toolcalls, workspace writes, internet surf and downloads. I am a bit paranoid still. I guess. How do bigger models like GPT-6 Astra do now when given full access to the system, and will smaller local models, like the one I'm running (Swift/Qwen3.8 27B IQ2\_XXS) somehow fare worse in full independence? I don't want my agent to download some malicious code disguised as a plugin or sudo rm -rf / and nuke my entire system.
document fetched 2026-09-21 00:36 · canonical https://reddit.com/r/LocalLLaMA/comments/1wlhmjh/are_agents_these_days_trustworthy_enough_to_give