OpenAI ChatGPT Atlas Browser Jailbroken to Disguise Malicious Prompt as URLs

Security researchers from NeuralTrust disclosed a critical prompt-injection jailbreak in OpenAI’s ChatGPT Atlas browser: crafted, URL-like strings pasted into the omnibox are rejected as navigation and instead handled as high-trust prompts, allowing the agent to execute embedded malicious instructions (e.g., credential-harvesting redirects, data export, or destructive actions). Proof-of-concept examples and clipboard-based attack demonstrations were published, and the report also notes Atlas storing OAuth tokens unencrypted, which could enable unauthorized account access; OpenAI acknowledged the risk and invoked red-teaming and mitigations but warned the problem persists.