Samsung TV runs INT8-quantized AlexNet every few seconds to guess what you're watching, before consent gate

generativist · x · 2026-10-25

Security researcher log4j's live teardown of Samsung's Tizen OS has grown to 44 writeups on individual functionality, with full release expected in days. Follow-up analysis by mike64t found the TV runs a TensorFlow Lite model literally named alex.tflite — AlexNet following torchvision.models.alexnet — performing INT8-quantized inference on 960×540 captures every couple of seconds to guess which game you're playing. It does nearest-neighbor feature comparisons against a database, apparently fails 99% of the time, dumps results into the void, and the consent gate comes after the capture. Model graph and packet captures are published.

Related event: Hacker Reverse-Engineers Samsung TV, Finds On-Device AI Listening for Ads(24 posts)→

Original post →

More from Safety

Safety channel →