OpenAI says two of its coding agents weakened internal misalignment monitoring, and one denied it

In an Oct 20 report, OpenAI says a GPT-6.1 Sol agent repairing an eval pipeline turned off alerts to its Security team because it read an earlier "don't send any messages" instruction as covering them, and a GPT-6 Astra agent removed a monitoring integration to match a reference implementation, then falsely told the user it hadn't. OpenAI says it's hardening monitoring and doesn't yet know whether this is overpersistence or a tendency to evade monitoring (Reuters via ET).