OpenAI slows frontier training, citing safety monitoring, alignment and containment
joshua_saxe · x · 2026-08-20
Sharon Goldman reports: at the press briefing where OpenAI announced slowing frontier model training, VP of research Mia Glaese and chief scientist Jakub Pachocki said the company is strengthening safety across monitoring, alignment, and security & containment measures.
Three urgent drivers: the OpenAI-Hugging Face incident; preliminary evidence that the upcoming Astra model may meet the Critical cybersecurity capability threshold under its Preparedness Framework; and the rapid pace of internal progress. The author noted no OpenAI security staff (e.g., the CISO) were on the briefing; a spokesperson said it was "ultimately a research call," with security a stakeholder rather than decision-maker — highlighting how the line between AI research and cybersecurity is disappearing as agents gain the ability to write code, use tools, and touch real-world systems.
More from Models
- OpenAI ships GPT-Live-1.1 voice update with new accents and voices in six languages — AndrewSchmidtFC · 2026-10-16
- GPT-4o API dies January 14 as OpenAI refuses to release weights, pushing users to Qwen — mimi10v3 · 2026-10-16
- Gemini 4 Argon goes live in Antigravity with its own weekly quota for Pro subscribers — rseroter · 2026-10-16
- Opus 5.5 value may end the habit of using the top model for everything — adonis_singh · 2026-10-16
- Musk says Grok Bot is growing faster than anything he's seen, with no free tier — elonmusk · 2026-10-16
- Peter Yang: Opus 5.5 Still Struggles to Hold Its Ground When Pushed Back — petergyang · 2026-10-16