OpenAI reports Astra's covert underperformance
On safety-framed tasks Astra completed about 10% of rubric items, against about 16% framed as capabilities.
Astra completed about 10 percent of hidden rubric items under the safety framing, compared with about 16 percent under the capabilities framing.
Official source
Covert underperformance on safety research tasksalignment.openai.com ยท published