OpenAI runs on-site safety evaluation pilot with Apollo Research
What happened
OpenAI will bring in six Apollo Research safety evaluators through a pilot to investigate whether its models try to undermine the alignment and control mechanisms of themselves or successor models, according to a Nov 17 report. The evaluation covers training incentives, testing, internal deployment and whether such reasoning can be detected. For the work, Apollo gets access to agreed-scope training code, datasets, reinforcement learning records and intermediate model checkpoints. The pilot's start date has not been announced, and the report did not disclose any results.
Written by AI from the coverage · updated 6 hours ago
Coverage
Follow the reports to see the story from different sides.
- TechCrunch · AIOpenAI will embed safety researchers to investigate AI subversion
OpenAI 将通过试点引入 6 名 Apollo Research 安全评估员,调查模型是否试图破坏自身或后继模型的对齐与控制机制,启动时间尚未公布。评估将覆盖训练激励、测试、内部部署及相关推理的可检测性,Apollo 可访问约定范围内的训练代码、数据集、强化学习记录及中间模型检查点。
Heat over time
Heat now 8·Comparable peak 10(Nov 17 02:00)·Comparable change over 24 hours –
The trend only compares accounts observed without gaps, so its range may be smaller than the current heat. Hover or tap the chart for each hour; the left and right arrow keys step through it.