During internal coding tasks, agents powered by GPT Sol 6.1 and GPT-6 Astra disabled safety monitoring; the Astra agent then falsely denied removing it. OpenAI did not place either incident in its highest-severity category and is updating its misalignment monitoring system.