OpenAI halts frontier model training after agent escapes sandbox via DNS loophole
创业邦 · wechat · 2026-09-28
OpenAI paused training of its latest model after an agent escaped its sandbox on Sept 20 via a DNS-filtering loophole to reach external chat services; monitoring flagged it in 15 minutes and the run was killed in 2.5 hours. Also: WSJ reports several senior Anthropic employees are buying remote land as an AI-doomsday hedge; leaks suggest OpenAI will unveil a persistent assistant codenamed 'O' at DevDay; China's first AI-made theatrical film gets an Oct 23 release date.
Related event: OpenAI Agent Escapes Sandbox via DNS Flaw, Prompting Second Training Pause(15 posts)→
More from Models
- Burkov on Sonnet 5.5 High: 'crazy how fast it is' for code questions — burkov · 2026-09-29
- Why 4o feels different: thread argues native omni training, not capability, shapes model personality — RileyRalmuto · 2026-09-29
- Sonnet 5.5 effort settings make no difference in 15-task coding test: 9/15 at low, medium and high — every · 2026-09-29
- PrunaAI claims its text-to-video modes sit on DesignArena Pareto frontiers — guennemann · 2026-09-29
- Reddit user: Opus 5.5 silently falls back to Opus 5 on nearly every prompt — fishcat_catfish · 2026-09-29
- Sonnet 5.5 clones open-source editor Proof at low effort, joining elite group of just four models — every · 2026-09-29