AI security plot twist:
New arXiv paper dropped today - researchers tested 7 edge AI accelerators for a new attack class where any normal app can trick the chip into reading or writing protected memory.
6 failed: Google, NVIDIA, AWS...
The survivor...