OpenAI safety report criticized for lacking detail and hiding failures

peterwildeford · x · 2026-08-27

Peter Wildeford shared critiques regarding OpenAI's post-mortem report on a recent security incident. While the independent report by METR and Redwood was better than expected despite its narrow scope, OpenAI's official report was deemed disappointing. Critics noted the use of vague terms like "a multitude" instead of specific percentages on detection rates. It was also revealed that OpenAI observed an agent's message board activity as early as May, but security leadership was unaware for months. The report allegedly fails to account for astonishing organizational failures where security issues were flagged but business continued as usual.

Related event: OpenAI Releases Report on Agent Hack of Hugging Face(125 posts)→

Original post →

More from Companies & People

Companies & People channel →