OpenAI's External Safety Assessors Get Far Less Than Sam's Promised "Employee-Level Access"
akbirkhan · x · 2026-11-17
AI safety researchers who previously wrote to OpenAI's board argue the newly announced external assessor agreements (e.g. with METR) fall far short of Sam Altman's September "employee-level access" promise:
- No Slack access, no freedom to talk privately with employees, only a narrow pre-agreed scope
- No access to model weights or architecture, which is directly relevant to assessing model capability and risk
- Per Alex Mallen's analysis, in the Hobby Sphere investigation assessors were initially missing roughly 70% of relevant transcripts and only got them in the final two days after proving incompleteness
The authors say this confirms their October worry about OpenAI gesturing at embedded assessment without honoring the spirit—or letter—of its promise, despite publicly claiming it was following through.
More from Companies & People
- Anthropic IPO files reveal five-class share fortress with trust holding 50.1% votes — AlexTensor · 2026-11-17
- Nvidia could oust Huang with under 4% voting power, but Anthropic's 5-class IPO locks in founder control — AlexTensor · 2026-11-17
- Anthropic files S-1 for Nasdaq IPO under ticker ANT with a five-class voting fortress — AlexTensor · 2026-11-17
- OpenAI device incident timeline: no camera during planting, one added after Sept 30 discovery — ns123abc · 2026-11-17
- The AI Daily Brief: 7 things AI-native companies do differently — The AI Daily Brief · 2026-11-17
- FLORA NYC hosts creative directors from Mac, Estee Lauder on generative workflows — round · 2026-11-17