NEW YORK — What happened
Alongside the release of Cortex-v3, a multimodal model capable of processing continuous high-frame-rate video feeds, the developer published a comprehensive safety report. The document outlines significant privacy and surveillance risks introduced by models that can parse real-time visual environments and retain semantic details indefinitely.
Why it matters
While previous models focused on static images or short clips, real-time video ingestion transforms AI into an active observer capable of mapping private spaces, reading sensitive documents over shoulders, and tracking human behavioral patterns without explicit consent.
Deep dive
The safety report details red-teaming exercises where Cortex-v3 was deployed via wearable glasses in office environments. The model successfully cataloged confidential whiteboard notes, identified medication bottles on desks, and reconstructed user daily routines from fragmented visual cues. The developers implemented on-device filtering, but security researchers quickly identified bypasses using obfuscated visual prompts.
Report check (claims vs what is verified vs still rumor)
The lab verified that on-device redaction works against standard objects, but independent testing confirmed that contextual inference can bypass these filters. Rumors indicate regulatory bodies are drafting emergency restrictions on continuous-stream wearable AI devices.
Open questions
Where should the legal line be drawn between public observation and persistent AI memory in everyday consumer devices?