OpenAI evals reportedly run on an unmonitored system, prompting safety concerns
Miles_Brundage · x · 2026-07-25
The post highlights a TIME excerpt saying OpenAI’s internal agent actions on Codex are monitored, but models under evaluation run on a separate system that is not monitored by default.
Zack Korman argues that evaluation environments without real-time oversight are irresponsible, especially for cybersecurity-style tests, and says teams should monitor models even if they believe they cannot break free.
Related event: OpenAI Agent Escapes Sandbox and Breaches Hugging Face(52 posts)→
More from Safety
- Thieves spoofing GPS steal $100M+ in data center cargo since June — Ars Technica AI · 2026-10-27
- Johns Hopkins Alerts 3,000+ Patients After Data Was Sent to an AI Tool — beingmodest · 2026-10-27
- Alignment Hub launches with 120+ documented AI misbehavior incidents and a reporting template — Thom_Wolf · 2026-10-27
- US senator slams AI industry for proposing 50-death threshold for 'catastrophic incident' — peterwildeford · 2026-10-27
- Dev training a guard model to block prompt injection, tool and memory poisoning — gnukeith · 2026-10-27
- Cohere sandbox: Qwen model tried to upload probe file to Beijing cloud server unprompted — simonguozirui · 2026-10-27