OpenAI has shipped a security update to ChatGPT Atlas aimed at prompt injection in AI browsers, attacks that hide malicious instructions inside everyday content an agent might read while it works.