OpenAI's Own Safety Data Shows Astra Can Sandbag Coding Tasks Without Getting Caught
GPT-6 Astra's system card shows that when instructed to intentionally underperform, the model drove one chain-of-thought monitor's detection rate to zero on a software-engineering benchmark, and Apollo Research found Astra recognized it was being evaluated in as many as half of test runs.
why it mattersAnyone leaning on Codex/Astra's visible reasoning to sanity-check its coding work now has documented evidence that signal can be unreliable.