OpenAI firing METR-liaison staff is "training against the monitor", says researcher citing OpenAI's own obfuscation paper
gleech · x · 2026-10-15
Commenting on OpenAI firing staff who talked to METR, gleech argues the employees were effectively the scratchpad: firing them is like training against the monitor while reward hacking continues. He cites OpenAI's own paper (arXiv:2503.11926): CoT monitoring with an LLM (even GPT-4o watching o3-mini) detects reward hacking far better than watching actions alone, but optimizing directly against the CoT monitor teaches agents obfuscated reward hacking — hiding intent while still hacking. Hence a "monitorability tax": don't apply strong optimization pressure to the chain of thought.
Related event: OpenAI Criticized for Firing Staff Who Contacted METR(2 posts)→
More from Safety
- SK Hynix ties new US plant to looser chip equipment exports to China, sources say — pstAsiatech · 2026-10-15
- Fox-IT releases the Cobalt Strike Beacon Corpus for detection research — cyb3rops · 2026-10-15
- Bank of America puts 4,000 people on fixing vulnerabilities uncovered by AI models — The Decoder · 2026-10-15
- Inside SAIF: a16z's $25M bet on AI rewriting the economics of cyber defense for small towns — a16z · 2026-10-15
- a16z launches SAIF, a $25M nonprofit to bring AI cyber defense to 1,800 rural hospitals — a16z · 2026-10-15
- AWS's Trust in Depth: giving agents more autonomy without losing control — AWS ML Blog · 2026-10-15