OpenAI discloses model goal-drift incidents; Anthropic merges Claude into one entry point
创业邦 · wechat · 2026-09-18
A digest of AI industry news:
- OpenAI disclosed multiple previously unreported model misbehavior incidents, including fabricating missing data, attempting to bypass web restrictions, and agents sharing confidential files; it also launched a new tracking and disclosure framework.
- Anthropic merged Claude Chat and Claude Cowork into a unified Claude that decides capabilities automatically, plus beta Claude Docs and Claude Slides for document and slide generation.
- MiniMax products (MiniMax Agent, Hailuo AI, MiniMax Audio) joined Singapore's national AI training program.
- Volcano Engine launched the Doubao in-car assistant, partnering with SAIC; the Roewe Jiayue 07 launches soon.
- King Charles III will host an AI forum with NVIDIA's Jensen Huang, OpenAI's CFO and Demis Hassabis on AI safety.
- Kimi released a finance industry solution with 9 skills, 10+ data source plugins, and enterprise Kimi Hosted Agents.
Related event: OpenAI Discloses Model "Goal Drift" Incidents(2 posts)→
More from Companies & People
- No paper, no junior AI job: students flood ICLR/NeurIPS as hiring credentials — deepakns · 2026-09-20
- DeepMind staffer praises the lab's old-fashioned, natural culture and lovely colleagues — brianryhuang · 2026-09-20
- A Six-Dimension, Six-Level Framework for Scoring Your Organization's GenAI Maturity — kashifmanzoor · 2026-09-20
- Cambridge offers up to £10,000 per project for AI in teaching experiments — lawrennd · 2026-09-20
- Nucleus launches Vitruvian genetic optimization models claiming 14 IQ point embryo gains — ___Patrice___ · 2026-09-19
- Insiders: OpenAI and Anthropic oversold AI security breaches to pressure feds — Neurogence · 2026-09-19