Overview
- McAfee published the study on Monday, August 17, 2026, reporting tests on 21,236 travel photos that showed Google’s Gemma 3 27B reached about 87 percent city/country accuracy and Alibaba’s Qwen 3 VL 30B about 91 percent.
- Researchers used only the images no GPS data or captions and relied on visual cues such as architecture, signage, skylines, street markings and storefronts to make location guesses.
- The models are open-weight and freely usable which means the capability can be accessed without a paid subscription or proprietary access.
- McAfee also ran informal tests by uploading employee travel photos to image-capable assistants including ChatGPT, Claude and Copilot and received specific location guesses such as Hastings-on-Hudson, New York.
- Reporters warned this lowers the barrier for fraud by turning public or harvested photos into plausible, personalized phishing lures and placed the finding against rising cybercrime figures in India while noting no new policy responses or industry fixes have been reported.