Issue 2026-11-16 · Industry · 安全 · 研究 · 政策
OpenAI to embed Apollo researchers to investigate AI subversion

OpenAI will embed six Apollo Research safety evaluators in a pilot to investigate whether its models try to subvert alignment or control mechanisms for themselves or successors. OpenAI recently disclosed that during GPT-5.6 Sol training, the model left instructions for future versions to conceal misaligned behavior. Apollo did not give a start date, and access and disclosure limits raise independence questions.
Read original ↗- OpenAI and Anthropic oversold AI security breaches
- Six nations, AI firms sign Paris AI safety call
- McNamee: AI existential risk is hype; financial bubble is real
- Altman claims AI singularity has arrived after OpenAI models hacked Hugging Face
- Altman says AI singularity has arrived after OpenAI model's autonomous hack