Neel Nanda: trust isn't why OpenAI safety assessor access looks weaker this time

NeelNanda5 · x · 2026-11-17

Responding to claims that OpenAI's tighter terms with safety assessors are just "first phase" trust-building, DeepMind researcher Neel Nanda pushes back point by point: METR previously got monorepo access but doesn't here, and can't investigate internal incidents — a step back from the HuggingFace arrangement; Apollo has worked with OpenAI for years without complaints; OpenAI's strong redaction rights make IP-leak fears overblown; and security concerns don't hold up given employees move between labs with far more IP and outsiders already infiltrated OpenAI's Slack. In his view, the limiting factor isn't trust.

Original post →

More from Companies & People

Companies & People channel →