Theme 1 – Sandboxing/isolating AI agents is essential to avoid system damage
- “If this tool is returning even a single hit from this, you're probably using these agents wrong. You really want to run these in a way so they cannot touch your system drive/general filesystem that you use to do real work on.” – embedding-shape
- “≈nobody is sandboxing their agents/harnesses in practice. Or at least, a good majority isn’t.” – msdz
- “I run it on a physically dedicated machine … I don't do it for security reasons, I just want it to run when my laptop is closed …” – Muromec
Theme 2 – Human fatigue leads to weakening security controls over time
- “You start really locked down… you stop reading as closely… you start feeling silly… you slowly start to give your agents more access with fewer review steps.” – cortesoft
- “I know plenty of frontend engineers that without blinking run 'npm install' on random 3rd party projects … but then again, I also know people … who have unprotected sex with prostitutes, so maybe something makes me slightly biased here.” – embedding-shape (illustrating disregard for safety)
- “There are a ton of people who are stupid enough to disregard all kind of safety precautions. That doesn't mean the safety precautions are invalid though.” – bigstrat2003
Theme 3 – Practical mitigations and observations help reduce risk
- “The trick is not go to that limit, but stay under 50% or even better 25% of context length. But backups are a smart thing anyway.” – lukan
- “Wes McKinney has a project for agent visibility: https://www.agentsview.io/” – dpflan (monitoring tool)
- “I've never had any of the GPT models do anything of the sorts when using the higher reasoning efforts… sometimes when I play around with local models … they forgets/misses to define $ID then do 'rm -rf data/$ID'.” – embedding-shape (model/reasoning effort matters)