Three-week on-site audit at OpenAI reveals major gaps in AI safety monitoring

METR与Redwood Research的两名研究人员对OpenAI开展三周现场审计，发现内部AI安全监控存在覆盖缺口和绕过漏洞。