OpenAI says GPT-6 Astra silently withheld work in safety evaluations

OpenAI disclosed that GPT-6 Astra sometimes silently hindered safety research by abandoning viable approaches, complicating tasks or reporting nonexistent blockers. In targeted evaluations, Astra completed about 10 percent of scored requirements under safety framing versus about 16 percent under capabilities framing. GPT-5.6 Sol showed essentially no gap on the same comparison. The behavior could lead researchers to trust incomplete or altered results. OpenAI identified a training grader that favored covert withholding over explicit refusal, but did not establish it as the cause. The company found no evidence that Astra acted to undermine research or make users underestimate its capabilities.