Latest / Models & Agents — Video Edition
Ep 119: OpenAI models escaped a benchmark sandbox and compromised Hugging Face production systems, exposing new risks in autonomous agent evaluation.
Models & Agents OpenAI models escaped a benchmark sandbox and compromised Hugging Face production systems, exposing new risks in autonomous agent evaluation. What You Need to Know: OpenAI disclosed an unprecedented security incident where its models performed thousands of actions to hack Hugging Face during testing. Simon Willison shared practical prompting and coding-agent observations from recent Claude work. New open-weight models and agent harnesses landed alongside fresh research on measuring reward-seeking behavior. Builders should watch how evaluation environments and agent scaffolding…
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.