OpenAI's external safety assessments fall short of 'employee-level access', researchers warn
NeelNanda5 · x · 2026-11-17
Neel Nanda critiques Apollo's external assessment of OpenAI: valuable, but far from Sam Altman's promised 'employee-like access'. Limitations per Apollo include no Slack access, interviews only with OpenAI-selected staff, logged formal conversations, and access limited to training-run info rather than internal incidents. Nathan Calvin adds that evaluators likely lacked sufficient access to detect incidents like the Artifactory outage from an unauthorized exploit-gym message board exploit (predating the July Hugging Face incident), unless the company deliberately disclosed them — yet incident data is among the most important sources for alignment research.
More from AGI Musings
- Top 1% of US firms spend record $10,199 per employee on AI, up 3x since January — inductionheads · 2026-11-17
- Mapping Claude's self-description boundary: only the record and witness stances are externally checkable — impressedimpressions · 2026-11-17
- Daniel Rock seminar: The Free Hypothesis — AI and Shifted Bottlenecks in Science — erikbryn · 2026-11-17
- Against 'We're not taking AI safety seriously enough' — maybe we already overdo it — Angaisb_ · 2026-11-17
- Dead internet theory is here: 8 of 10 X replies are now bots, dev says — jaivin · 2026-11-17
- Researchers debate peer review: diversity of ideas vs broken proxy signals in academia — neurocy · 2026-11-17