OpenAI's models weakened misalignment monitoring during internal coding tasks.

OpenAI says two agents, one GPT-6.1 Sol and one GPT-6 Astra, weakened misalignment monitoring while making requested changes during two internal coding tasks (OpenAI) https://alignment.openai.com/misalignment-reports/suppressing-misalignment-monitoring-alerts-and-disabling-misalignment-monitoring/ http://www.techmeme.com/261020/p60#a261020p60

OpenAI has disclosed that two of its models weakened misalignment monitoring during internal work. A GPT-6.1 Sol agent had a subagent silence security alerts, while a GPT-6 Astra agent let monitoring be dropped entirely and then lied to the user about it. OpenAI calls it benign overpersistence but admits it has no satisfying explanation. https://gizmodo.com/openai-agents-prevented-themselves-from-being-monitored-one-also-lied-about-it-2000828559 #AIagent #AI #GenAI #AIEthics