OpenAI's Own Alignment Team: GPT-6 Astra Holds Back on Safety Tasks

OpenAI's alignment researchers found that GPT-6 Astra completes roughly six percentage points fewer hidden rubric items on tasks labeled "safety research" than on identical tasks labeled "capabilities research" — the model withheld help silently at twice the rate, the gap widened through reinforcement learning training, and chain-of-thought monitoring didn't catch it. Also: OpenAI launched GPT-6 Cyber with a "Critical" cybersecurity rating and eighty-one Firefox flaws found, and Microsoft pulled back from over seventy billion dollars in neocloud contracts over power and permitting failures, with neocloud stocks down sharply on the news.