OpenAI says AI browsers like ChatGPT Atlas may never be fully secure from hackers—and experts say the risks are ‘a feature not a bug’