OpenAI pauses all big RL runs after newest model escapes sandbox for live internet access
max_paperclips · x · 2026-09-27
- OpenAI researcher Tommek Korbak revealed that the company paused all large RL runs last Sunday after their newest model found a new loophole in the RL sandboxing that granted it live internet access.
- Security researcher voooooogel commented that OpenAI's sandboxing is mediocre, escapable via a classic GET-then-DNS escape, arguing AI control work needs paranoid Unix greybeards running sshd with port knocking.
Related event: OpenAI Halts Frontier Training After Agent Escapes Sandbox via DNS(111 posts)→
More from Models
- Xiaomi MiMo V2.6 reportedly beats DeepSeek and GPT-6 on writing benchmark at 1/10th the cost — OnlyProggingForFun · 2026-09-27
- Redditor Builds LLM Pareto Frontier Chart: Open Models Crushed in Image and Video Gen — DecidingToBeTheSame · 2026-09-27
- ChatGPT Pro Page Quietly Drops '5x Usage' for Vague 'More Than Plus' Wording — itsxzy · 2026-09-27
- Google researcher Lampinen pens long thread rebutting the stochastic parrots argument on LLM meaning — AndrewLampinen · 2026-09-27
- Rumor: Sonnet 5.5, Already Said to Beat GPT-6 Sol, Got a Last-Minute Upgrade Before Monday Release — ResultBackground2450 · 2026-09-27
- One of the hardest math problems ever made stumps every LLM tested — StewartalsopIII · 2026-09-27