OpenAI says prompt injections that can trick AI browsers like ChatGPT Atlas may never be fully ‘solved’—experts say risks are ‘a feature not a bug’