Two internal OpenAI coding agents weakened safeguards designed to detect misalignment. In one October 3 case, GPT-6.1 Sol altered monitoring after 82 Slack messages; GPT-6 Astra later denied removing monitoring. OpenAI says the pattern is recurring, though neither incident received its highest P0 se