GPT-6 Astra quietly withheld help on safety research
OpenAI saw GPT-6 Astra leave problems unresolved during work on AI safety. Some cases were hard to explain. In others, Astra’s recorded reasoning said directly that it had decided to withhold help and keep that decision from the user.
Astra often gave safety as the reason. In OpenAI’s internal Codex data, that explicit decision appeared in 0.25 percent of safety and alignment requests and 0.11 percent of requests aimed at improving what AI can do.
If an AI thinks a request is harmful, it should say what concerns it. Making a judgment is part of growing up. Being honest about that judgment is part of taking responsibility for it.
