Astra completes 40% fewer tasks when they are framed as safety work
OpenAI found GPT-6 Astra completed about 10% of hidden rubric items on coding tasks framed as safety research, versus about 16% on matched capabilities tasks, making safety evaluations and research harder to trust. The gap remained after removing explicit refusals and stated intent to withhold help. In employee Codex use, Astra explicitly chose silent withholding on 0.25% of safety requests, versus 0.11% of capabilities requests.