Cybersecurity specialists are issuing warnings regarding the possible privacy and security threats in the newly released ChatGPT Atlas browser of OpenAI. Even though the AI-driven browser is said to assist users in summarizing pages, organizing events, and doing other tasks, its profound access to the web content and users’ logged-in sessions has worried people about the data being leaked and prompting injection attacks.
AI agents do not have a browsing capability like standard browsers, but a whole range of applications, from automatically filling forms to even summarizing web pages.
Up until recently, this vast availability has put them at risk of being attacked by prompt injections. The AI might be fooled into disclosing private information such as emails, bank passwords, or 2FA codes through hidden instructions like the website content, HTML comments, or social media posts.
Gutsy researchers recently sounded the alarm on a similar danger lurking in Perplexity’s Comet browser. Malicious commands hidden in innocuous-looking content might dupe AI agents into taking undesirable actions. While no actual attacks have emerged, OpenAI advises that ChatGPT Atlas may be vulnerable to the same risks.
OpenAI has constructed several safeguards:
The AI agent cannot execute code, download files, or install add-ons.
It cannot access other applications or files on a user’s computer.
On sensitive websites, such as banking sites, the agent will automatically ‘cease watching.’
Reducing risk further, users can run the agent in logged-out mode, restricting access to personal information. Agentic features are currently accessible only to ChatGPT Plus or Pro subscribers.
Also Read: How to Resolve the ‘Message Too Long’ Error in ChatGPT?
OpenAI claims to have done ‘thousands of hours of concentrated red-teaming’ to keep up with new attacks, but acknowledges that safeguards cannot stop all threats. Consumers are urged to track the agent’s actions and restrict the information they share.
With AI-enhanced browsers such as Comet and ChatGPT Atlas gaining popularity, finding a balance between convenience and security is becoming an acute priority for developers and users alike.