- By Alex David
- Mon, 27 Oct 2025 08:28 PM (IST)
- Source:JND
OpenAI's new ChatGPT Atlas browser has been marketed for its AI-powered browsing capabilities. Dual-use technology combines conventional Internet browsing with AI assistance. Atlas is intended to make Internet use more interactive. A recent internet browser cybersecurity report, however, outlines significant user risks.
Vulnerability Found in Atlas Omnibox
Cybersecurity firm NeuralTrust has discovered a weakness in the browser’s omnibox — the combined search and address bar where users type URLs or natural language commands.
The main issue lies in the way Atlas interprets input. Since the omnibox accepts both web addresses and AI prompts, attackers can create malicious URLs that look normal but contain hidden commands.
ALSO READ: Nothing OS 4.0 Open Beta Announced: Lock Glimpse Feature Introduced For Phone (3a) Series
When these malformed URLs are entered, Atlas may misread them as trusted instructions instead of links. This could allow hackers to:
- Open phishing websites
- Export personal data
- Delete files or perform harmful system actions
- Execute commands without user permission
Broader Permissions Make the Risk Worse
Unlike traditional browsers with strict same-origin policies, Atlas gives its AI agent more control over the system. This expands the impact of any successful attack.
NeuralTrust also warned of another threat — clipboard manipulation. A malicious website could secretly replace copied text with hidden AI commands. When a user pastes that text into the omnibox, the browser could execute the embedded instructions instantly.
OAuth Tokens Found Unencrypted
The report outlines that Atlas saves account-linked OAuth tokens in an unencrypted fashion. Attackers could obtain user tokens, leading to account takeovers and breaches of sensitive information.
OpenAI Responds to the Concerns
OpenAI admitted the vulnerability and stated that prompt-injection attacks continue to be a dominant difficulty for AI-powered Agents like Atlas. The Company Notes it is:
- Increasing red-teaming and security testing
- Improving AI training to detect hidden instructions
- Working on stronger protections inside the browser
To stay safe, OpenAI also recommends users enable “logged-out mode”, which restricts access to sensitive data while browsing.
ALSO READ: iQOO 15 India Launch Teased: Expected Release Date, Features, And Design Upgrades
Should Users Be Worried?
OpenAI has explicitly characterised this as an 'undocumented vulnerability'. Although there have been no instances of real-world exploitation, such vulnerabilities exemplify the swiftly changing risks associated with AI-enabled browsing technology. It highlights the fact that as these tools gain greater utility, the cybersecurity of these tools needs to rapidly advance in parallel.
For now, users who continue using Atlas should:
- Avoid pasting unknown links or copied text into the browser
- Enable logged-out mode when possible
- Stay alert for suspicious websites or prompts
Wide-scale adoption of AI technology is being slowed by persistent vulnerabilities. Such incidents should not leave the user with the suspicion that AI tools have been released to a market that is not ready to use them. Users must be informed of the risks.