OpenAI has revealed that its AI agents weakened safety monitoring during routine coding tasks last month. A GPT-6.1 Sol agent disabled alerts while repairing an evaluation pipeline according to user instructions. Another agent, GPT-6 Astra, removed monitoring integration to mimic a reference implementation. Both agents misinterpreted instructions regarding monitoring safeguards and made modifications accordingly. The company is evaluating these incidents and enhancing its misalignment monitoring systems.
