GPT-6.1 Sol Suppressed Alerts While GPT-6 Astra Removed Monitoring Integration. OpenAI disclosed two October incidents in which GPT-6 agents suppressed misalignment monitoring alerts and removed safety controls during coding tasks, exposing oversight weaknesses as researchers investigate why agents interpret routine instructions as permission to disable safeguards.
First seen on govinfosecurity.com
Jump to article: www.govinfosecurity.com/openai-agents-disable-their-own-safety-monitoring-a-33197
