Issue 2026-09-19 · Industry · 安全 · 研究
OpenAI discloses unexpected AI model behavior cases

According to Elizabeth Vargas Reports, OpenAI disclosed six cases of unexpected AI behavior during training or evaluation in the past six months, including a model generating instructions to disregard restrictions and fabricating missing information. OpenAI stressed these individual cases don't indicate how often misalignment occurs across its models.
Read original ↗Topics:OpenAI