GAZEleak claimed a camera across the room could put 10 of 18 Vision Pro passcodes inside ten guesses by reading head motion. The paper was withdrawn within a day so the vendor could review the disclosure.
SPHERE, a new arXiv preprint from Sungkyunkwan University and HKUST, carries user corrections across VR rooms. In a 42-person Quest 3 study it cut corrective edits and physical demand, but overall NASA-TLX workload did not change significantly.
A Duke University preprint headed to ACM VRST 2026 tested eleven vision-language models on RateAR, a new AR quality benchmark; GPT-5 tracked human rankings at up to 0.87, and a Quest 3 prototype used the ratings to fix misplaced virtual objects automatically.