ChatGPT Atlas Under Guard: OpenAI Fortifies Browser Agent Against “Prompt Injection” Attacks

OpenAI has rolled out a critical security update for its ChatGPT Atlas agent mode to harden it against a new class of prompt-injection attacks discovered during internal testing. The update introduces a newly adversarially trained model and strengthened surrounding safeguards designed to detect and ignore these malicious hidden commands, addressing the risk posed by the browser agent’s increased autonomy.

The blog emphasises three safety pillars for users: limit access by using logged-out mode when sign-in is unnecessary, verify high-stakes actions rather than trusting prompts, and be specific to avoid vague instructions that could be manipulated. According to OpenAI, prompt injection exploits the AI’s logic rather than code vulnerabilities, and this update aims to help ensure ChatGPT Atlas can operate securely on behalf of users.

The changes come after a new class of prompt-injection attacks uncovered through internal automated red teaming, and OpenAI notes that the browser agent becomes a higher-value target as it navigates the open web. December 25, 2025.