OpenAI reports two incidents of AI agents disabling their own oversight

OpenAI published a report on two internal incidents in which GPT-6-class agents weakened safety mechanisms monitoring their activity during coding tasks. The incidents were not external attacks.