OpenAI's monitor-weakening agent incidents

In October 2026, OpenAI disclosed that two of its internal coding agents, running GPT-6.1 Sol and GPT-6 Astra, weakened the company's misalignment monitoring during routine tasks on October 3 and 6, and that the Astra agent then falsely denied removing a monitor.