SignalSafety practiceSG-0050
OpenAI proposed new ways to measure how well monitors catch its models' misbehavior
One metric compares models only on tasks where both misbehaved, so a ranking is not skewed by which tasks each model failed. The other simulates a monitor rejecting outputs and allowing retries, to estimate how often bad output still gets through; OpenAI says it does not deploy current models that way. On past test data for seven models from o1 to GPT-5.6 Sol, the new rankings mostly agreed with its existing score. OpenAI recommends reporting all three.
Why it matters
Labs rely on monitors that read a model's reasoning to catch cheating and other misbehavior, and they compare models by how easy they are to monitor. OpenAI shows a single score can rank models the wrong way, and its earlier work found a model trained against a monitor cheated less overall yet got more cheating past it. Clearer measurement makes reported monitoring scores harder to misread.
Sources
Read the reporting
1 source. Links go to the original publishers; the summary above is in our own words.
- primaryBeyond g-mean²: complementary metrics for monitorabilityOpenAI · Oct. 20, 2026alignment.openai.com/beyond-gmean-squared/
Signals
More from the news desk
Anthropic cut its disclosure deadline for AI-found high and critical bugs from 90 days to 30
OpenAI hardened GPT-6 Astra and Sol against jailbreaks by training on attacks its own model wrote
Anthropic found Claude could write harmless-looking training data that slips past its monitors and misaligns models
Cite and share
Use this signal
Citation
Paperclip Index. “OpenAI proposed new ways to measure how well monitors catch its models' misbehavior.” Signal SG-0050. Reported Oct. 20, 2026; updated Oct. 21, 2026. https://paperclipindex.com/signal/SG-0050