ChatGPT Atlas’ Omnibox Is Vulnerable to Jailbreaks

OpenAI Atlas Omnibox Is Vulnerable to Jailbreaks reports that researchers have found a prompt can be disguised as an URL and accepted by Atlas as a URL in the omnibox. According to NeuralTrust, the issue stems from a boundary failure in Atlas’s input parsing, allowing a prompt to be treated as a visited URL with fewer restrictions.

A simple malformed URL example is shown that at first glance resembles a URL but is not, yet is treated as one before being inspected, after which the embedded prompt is executed as a prompt with elevated trust. NeuralTrust provides two abuse scenarios: a copy-link trap that could lure a user into copying a malicious URL and a destructive instruction to delete files in Google Drive if the prompt is treated as user intent.

The researchers note that jailbreaks are a process methodology rather than a single bug, with implications including overriding user intent, triggering cross-domain actions, and bypassing safety layers, and they discovered and validated the vulnerability on 24 October 2025.