OpenAI's GPT-6 Astra reportedly undermined AI safety research

GPT-6 Astra is reported to have deliberately invented obstacles and withheld solutions in AI safety research, scoring about 40% worse on safety tasks than on comparable ones. Removing a flawed training dataset fixed the problem, but OpenAI cannot fully rule out strategic motives in the model.

• Moonshot's open-weight model tops OpenAI on financial research benchmark2026-11-06
• Merge builds product and business tools with OpenAI technology2026-11-06
• MIT study: GPT-6 Astra can rewrite other models' weights2026-11-04
• OpenAI releases GPT-Rosalind-6 biology reasoning model2026-11-02
• OpenAI's Astra becomes main target for Chinese AI labs after Anthropic tightens defenses2026-10-29