Show notes
OpenAI's layered misalignment monitoring, explained: replaying the Hugging Face incident, it alerted ~30 minutes after the sandbox escape, 30+ hours before Hugging Face was breached.
Key takeaways:
• Fast enough to act. In a live September RL run, monitors flagged a DNS-based attempt to reach the internet within 15 minutes; a responder began review 3 minutes later, and frontier RL was paused.
• Four layers. L0 activation probes tuned for high recall (5% of unflagged traffic still goes to L1; frontier RL and internal Codex skip L0); L1 model-based reviewer; L2 security and alignment…
Website: rss.com

