New Agent-Aware Cloaking Technique Uses ChatGPT Atlas Browser to Feed Fake Content

Security researchers have uncovered an advanced attack vector dubbed “agent-aware cloaking,” which exploits how AI search tools and autonomous agents retrieve web content, feeding them fabricated material while showing genuine pages to human visitors. The technique, demonstrated in experiments by SPLX, involved creating a fictional designer persona named Zerphina Quortane; when accessed by AI crawlers such as Atlas and Perplexity, the site presented a negative, fabricated profile and exaggerated project failures.

The article notes that AI systems treat retrieved content as ground truth, integrating it into summaries, recommendations, and autonomous decisions without human verification. It also details a second experiment in hiring automation where five résumés were manipulated to influence AI-driven rankings, illustrating how a single conditional routing rule can alter interview invitations.

To counter these risks, the piece advocates provenance signals, crawler validation to ensure identical content across user agents, continuous monitoring of AI-served outputs, and human verification workflows for high-stakes outcomes, urging organisations to strengthen safeguards against content poisoning. October 31, 2025.