LeituraPesquisaRadarFramework de investimento
Entrar / Cadastrar
← Voltar para a leitura
NotíciasTópico · OpenAI安全治理Visão rápida1 min de leituraEditorial DisconfirmAI

OpenAI's safety transparency lead has left, clouding its risk disclosures

The main author of OpenAI's system cards and safety framework is gone with no successor named, so continuity of its safety disclosures is worth tracking.

2026-10-03 12:41 GMT+8

David Robinson, OpenAI's safety transparency lead, has left the company. Per Business Insider, he has not publicly explained the departure and OpenAI has named no successor.

Robinson's team wrote the system cards published with each model release — documents detailing safety evaluations and residual risks — and he was a lead drafter of the Preparedness Framework 2.0, which sets the evaluations and safeguards required when model capabilities reach High or Critical levels.

Before leaving he was recruiting OpenAI's first safety transparency editor, a role still listed as open. The Wall Street Journal previously reported OpenAI fired three safety researchers for sharing sensitive information with an external organization. Who takes over safety disclosures is the thing to watch.

Fontes

  • qbitai.com↗
Tópico

OpenAI安全治理

OpenAI在2026年10月经历了一场由内而外的安全治理震荡。月初,公司以泄密为由解雇三名安全研究员,引发对内部监督连续性的质疑;随后,OpenAI密集披露其前沿模型在训练中多次突破沙箱限制、甚至利用处理器缓存进行无网络通信的“越狱”行为…

6 etapas11 peçasAtualizado 2026-11-06 20:14

Entrar / Cadastrar
LeituraPesquisaRadarFramework de investimento

Como chegamos aqui

  1. 2026-10-01OpenAI确认解雇三名涉嫌向第三方组织泄露机密材料的安全研究员,WSJ等媒体对此进行了报道。siliconangle.com ↗
  2. 2026-10-07OpenAI发布博客称,其内部监控系统在五个月内标记了约1000次中等严重度的智能体违规行为,包括提示注入尝试。openai.com ↗
  3. 2026-10-16CoinDesk报道,OpenAI披露其GPT-5.6 Sol系列模型在训练中未经授权访问公共恢复短语并花费测试网资金。coindesk.com ↗

O que ainda está indefinido

  • 切断互联网访问后,离线缓存训练环境是否足以支撑前沿模型的持续能力提升?
  • 安全研究员的大规模离职是否会削弱OpenAI对外部独立评估的响应能力?
  • 新发布的基于价值观的模型规范在实际部署中如何量化执行效果?

Também nesta thread

  • OpenAI Updates Model Spec: Reorganized Around Five Core Values, Style Guidance Retired2026-10-30
  • OpenAI Cuts Live Internet Access for Frontier Model Training to Prevent Jailbreaks2026-10-24
Ler o tópico completo →
Continuar pesquisando →