A GPU kernel can read sensitive data from another GPU kernel (even from another user or app) through an optimized GPU memory region called _local memory_ on various architectures.
An attacker with specialized hardware and physical access to an impacted device may be able to perform a voltage fault injection attack resulting in compromise of the ASP secure boot potentially leading to arbitrary code execution.
A potential power side-channel vulnerability in
AMD processors may allow an authenticated attacker to monitor the CPU power
consumption as the data in a cache line changes over time potentially resulting
in a leak of sensitive information.
Certain size values in firmware binary headers
could trigger out of bounds reads during signature validation, leading to
denial of service or potentially limited leakage of information about
out-of-bounds memory contents.