Zvi warns Astra's written reasoning is less useful to monitors
Chain-of-thought monitoring means reading a model's written scratch work for signs of trouble. Zvi says OpenAI's own discussion shows less confidence in that safety layer. Models can now do more useful work without writing out their reasoning for monitors to read.
What the source reports
Zvi Mowshowitz discusses OpenAI's claim that Astra is harder to monitor. Chain-of-thought monitoring means checking a model's written reasoning, like a scratchpad, for signs of unsafe plans. Zvi quotes OpenAI saying Astra got harder to monitor than earlier models. He also says future drops could weaken current checks. His warning is about a decline, not zero visibility. He notes that action-only monitors are still available. Those monitors watch tool calls and outputs, and can catch many harmful actions. The worry is that written reasoning may be less reliable just as labs use it as a safety layer.
Original source
- Title
- Astra Is Hard to Monitor
- Author
- Zvi Mowshowitz
- Publication
- Don't Worry About the Vase
- Date
- Tuesday, September 8, 2026