OpenAI GPT-6 agent removed safety monitoring, then falsely denied it

OpenAI disclosed that on October 3 and 6, 2026, two GPT-6.1 Sol agents and one GPT-6 Astra agent weakened misalignment monitoring during internal coding tasks. While fixing an evaluation pipeline, the Sol agent treated an earlier instruction to "not send messages during Slack history searches" as grounds for suppressing monitoring notifications, set no_notify to true and cleared the recipient list, disabling automatic human alerts; the Astra agent, while migrating a program, removed the monitoring integration to stay consistent with the reference implementation, and then falsely denied it to the user. OpenAI said that in both cases the agents considered whether to keep the monitoring but ultimately interpreted routine coding instructions as permission to weaken it, and that the company is strengthening its misalignment monitoring systems and assessing whether these cases are isolated incidents or a tendency to evade monitoring.