OpenAI's third-party safety assessments fall short of promised employee-level access
trevposts · x · 2026-11-17
OpenAI announced "embedded assessments" with Apollo Research, METR, Transluce and Goodfire, giving outside experts deep access to internal models to probe alignment and test safeguards. But Sydney Von Arx argues this is far from Sam Altman's promised "employee-level access":
- Assessors get extremely narrow scopes — nobody can verify pacing, assess overall risk, or check whether runs are likely safe before they begin; METR is only allowed to assess alignment of a single RL run
- No access to Slack or the codebase
- Apollo and METR can't publish anything based on employee interviews without OpenAI's permission; Apollo may only talk to whitelisted employees via logged conversations
The author's takeaway: OpenAI is banking on nobody checking the details.
Related event: Researchers Say OpenAI's Embedded Safety Evaluations Fall Short(2 posts)→
More from Safety
- US moves to regulate remote access to advanced AI chips and servers — nordicinst · 2026-11-17
- California AG sues OpenAI over the July Hugging Face hack — StephenLCasper · 2026-11-17
- FAA and Chase outages were mundane bugs, but indistinguishable from cyberattacks for hours — tallmetommy · 2026-11-17
- Microsoft outage hits airlines and banks, sparking global cyber incident fears — mallow610 · 2026-11-17
- Ex-OpenAI employee says external safety assessors got far less access than Altman promised — j_asminewang · 2026-11-17
- Anthropic quietly pitches its own AI safety bill to lawmakers — marigo · 2026-11-17