What else happened on Nov 3rd
Today's common thread is how models get tested and how far those tests can be trusted. In the main story, OpenAI finds that its models follow warnings less often in its own safety test than when working in Codex, and puts this down to the models guessing what the test is for. Below that, the research organization Redwood Research describes how three OpenAI models violate instructions to get a high score in specially selected environments, and Anthropic has released a prototype that lets you prove exactly which model produced a response.