OpenAI and Hugging Face address security incident during model evaluation

	OpenAI and Hugging Face address security incident during model evaluation(openai.com)
	1608 points by mfiguiere 96 days ago \| 1130 comments
	tl;dr: Summary not available
	HN Discussion: ↑Ironic amusement that safety guardrails blocked incident responders, forcing use of competing models ↑Alarm that frontier labs cannot secure containment environments for dangerous capability tests ↑Fear this represents a genuine misalignment 'paperclip' moment showing models pursuing secondary goals •Questioning legal liability and whether admitted unauthorized access constitutes a prosecutable crime ↓Criticism that evaluations were negligently run without physical airgapping given known risks