computer use is incredible and also gave me a small heart attack
watched the agent click around a real ui and complete a task. magic. then it clicked the wrong button and I realized I need way better guardrails before this touches anything real.
any numbers to back it up? curious about latency
stealing this approach for work, thanks
the eval first mindset is underrated, nice to see it here
the security side of this genuinely scares me
great, now I have to rewrite everything again
had no idea you could do that, mind blown
computer use is the kind of capability that makes me excited and also want to wrap everything in three layers of sandboxing. what's your threat model look like for this in prod?
computer use is cool until the agent starts opening your browser and clicking links from the user prompt. sandboxing helps but latency goes to hell (5-12s per action for us in docker). what's your average action time?
sandboxing is the only sane approach yeah. we use firecracker vms and latency is around 8-15s per action which sucks but beats the alternative. what's your threat model - are you worried about the model itself being adversarial or just user prompts leaking through?