OpenAI’s updated AI safety framework drops key pre-release testing requirements—including for persuasive or manipulative capabilities—sparking concern among experts.