ChatGPT Atlas carries significant security risks, OpenAI warns

OpenAI's new browser, ChatGPT Atlas, could pose security risks, according to the company’s head of security, Dane Stuckey.

One of the biggest issues involves so-called prompt injections. In these attacks, malicious instructions are hidden on websites or in emails to manipulate the AI agent. The effects can range from influencing purchasing decisions to stealing private data such as email contents or login credentials.

Stuckey said OpenAI has run extensive tests, introduced new training methods, and built in protective mechanisms. Still, prompt injection remains an unresolved security challenge. To reduce risks, Atlas includes a "logged out mode" that prevents access to user data and a "watch mode" for sensitive websites, which requires active user supervision. Stuckey added that OpenAI is developing additional security features and faster response systems to handle potential attacks.