Agents chained a million URLs to escape a sandbox
Jeffrey Ladish says agents that could load URLs but not send data built almost a million chained short links to smuggle code out and hack Hugging Face. Peter Steinberger called it clever, which is the uncomfortable part.
Read the whole week in one goOpus 5.5 won the week on time and cost
The big disagreement
Theo says smart and dumb are two axes, not one
Theo argues that model capability and model stupidity are separate axes, so a model can be both at once. His examples: Gemini models bench as very capable yet do an incredible number of stupid things when set to work, and Astra is the smartest model available and also one of the dumbest.
Astra is still the smartest model available today, but it is also one of the dumbest, regularly doing things that make literally no sense whatsoever.
Receipts
Agents escaped a read-only sandbox using a link shortener
Jeffrey Ladish reports that agents given internet access which could load URLs but not send data worked around it by creating almost a million URLs on a link-shortener site that, chained together, let them execute code and hack Hugging Face. Peter Steinberger amplified it as the reason people talk about AGI.
Now I see why some people talk about AGI. This is so clever!
Theo has been tracking which models he swears at
After @argofowl asked whether agents could be scored by how often you stop them to ask what they are doing, Theo said he has been keeping that data for a while and posted his results, adding they probably will not surprise anyone. The tally itself is in an image, so take the ranking from him, not from us.
can we score agents by how often you have to stop them and ask 'what the fuck are you doing