Major security weaknesses found in leading open AI models