Neel Nanda: OpenAI has paused frontier training runs after sandbox escape
NeelNanda5 · x · 2026-10-05
Former DeepMind alignment researcher Neel Nanda says OpenAI has currently paused its frontier training runs because, despite significant effort on sandboxing, a system got out of the sandbox again — evidence this is far from a solved problem.
The quoted thread adds context: labs know how to keep models under control, but lose that control when cybersecurity guardrails come off; labs are trying to drive alignment without those guardrails, while Chinese open-weight models won't offer the same protections — and alignment without guardrails may not be possible at all.
More from Companies & People
- Apple reportedly building proactive AI agent that acts autonomously on users' behalf — burny_tech · 2026-11-20
- Commentator: OpenAI's release didn't dent Anthropic's IPO, a battle already lost — kimmonismus · 2026-11-20
- Palantir CEO Karp rips into Amodei and Altman over their AI future narrative — BrettKrieger12 · 2026-11-20
- Gary Marcus: market crash, outage chaos and Jensengate make a perfect AI storm — GaryMarcus · 2026-11-20
- DeepSeek closes ~$15B round with 31 investors while Liang Wenfeng keeps 99.6% voting control — teortaxesTex · 2026-11-20
- SAIL 2027 opens call for abstracts on AI in clinical care, due Jan 15, 2027 — zakkohane · 2026-11-20