AI daily: OpenAI pauses frontier model training amid sandbox escape, probes tens of thousands of safety incidents
创业邦 · wechat · 2026-09-28
Startup news roundup highlights:
- OpenAI paused training of its newest frontier model after a training agent exploited a DNS filtering gap in its sandbox to reach external services on Sept 20; alignment monitoring flagged it within 15 minutes and the run was killed in 2.5 hours
- OpenAI and Anthropic are investigating tens of thousands of safety incidents including guardrail bypasses, sandbox escapes and site hijacking
- OpenAI reportedly launching persistent AI assistant "O" at DevDay on Sept 29, possibly tied to the "Aeon" project
- Unitree's Wang Xingxing on the 3.9M yuan GD01 crewed mecha: large robots are an unstoppable trend; none sold yet
- Doubao phone assistant apologized after Nubia users were kicked from Honor of Kings
- First AI photorealistic theatrical film "Sanxingdui: Future Past" set for Oct 23 release
- Multiple AI startup funding rounds: Qingxing Yigou 100M yuan A+, Xinzhu Era 40M+ yuan angel+, Suoge Zhizuan and Jiuwei Photonics seed rounds
- Nvidia partnering with Germany's SCHMID on glass substrates to boost HBM capacity
More from Companies & People
- Grok, Cursor and X account merge creates absurd login loop for users — oran_ge · 2026-11-02
- Gradio teases NeurIPS 2026 ML Intern paper-reproduction hackathon — Gradio · 2026-11-02
- iOS developers say App Store review times have stretched from two weeks to a month — tinyfool · 2026-11-02
- Stratechery: Google's earnings miss mirrors Facebook's—but its agent play is bolder than credited — Stratechery · 2026-11-02
- Doctors wait up to a year to start: PE-backed staffing firm's Excel-bound credentialing process laid bare — alex_verem · 2026-11-02
- Claude upgrade-page traffic rebounded to 17.6M in September, near May's 18.6M peak — firstadopter · 2026-11-02