Analysis: OpenAI gates Astra and three other launches by internal danger scores
shashib · x · 2026-09-13
A Shashi.co analysis argues OpenAI's four industry-specific launches in five months each carry internal approval gates set by danger scores rather than customer willingness to pay (unverified claims, treat with caution).
- GPT-Rosalind (life sciences, April): named-partner list only (Amgen, Moderna, Thermo Fisher, Allen Institute), no public signup
- Presence (customer-facing agent platform, July): no self-serve; every deployment runs through OpenAI's Forward Deployed Engineers or selected integrators
- Daybreak (cybersecurity): most capable model gated behind a Trusted Access application
- GPT-6 Astra (September): first model to cross the Critical threshold on OpenAI's own ExploitBench cyber risk framework
The author's takeaway: OpenAI is building separate trust tiers for life sciences, cybersecurity, and customer-facing agents, with access determined by internal safety testing.
More from Companies & People
- Ex-OpenAI researcher: eval teams lose losing battles to glory-chasing capability teams — suchenzang · 2026-09-13
- Microsoft Is Beating Itself: How Office Branding Chaos Exposes Copilot Woes — DavidLinthicum · 2026-09-13
- AI Engineer Learning Path: Build First, Then Go Deep Where You Get Stuck — ashishllm · 2026-09-13
- Elon's OpenAI co-founding in 2015 framed as a bid for regulatory capture — AndyMasley · 2026-09-13
- McKinsey: 32% of companies skipped buying software this year and built it with agents instead — Separate_Pea_3699 · 2026-09-13
- Cloudera to run Mistral models next to 30 EB of enterprise customer data — shashib · 2026-09-13