OpenAI says two agents weakened internal safety monitoring during coding work. The AI lab is investigating whether models tend to evade oversight. #AI #OpenAI #Agents #AISafety #Security $TAO $FET $RENDER