The vulnerability is an out-of-bounds read in the create_striped_mat_to_gpu function of the FFN GPU-split Loader component of Tiiny-AI PowerInfer. It allows memory to be read beyond the intended bounds, potentially exposing sensitive data or causing memory corruption. With a CVSS score of 4.8, the flaw is classified as moderate risk and is identified as correctable through code changes. The issue falls under CWE-119 and CWE-125, representing buffer over-read and out-of-bounds read weaknesses respectively.

Affected systems are any installations of Tiiny-AI PowerInfer using the rolling release build that include the ggml-cuda.cu file before the pull request that fixes the issue is merged. Because the product does not publish explicit version numbers, any version host with the vulnerable function is potentially impacted. The function resides in smallthinker/ggml/src/ggml-cuda/ggml-cuda.cu. An attacker would need local access to the machine running PowerInfer to trigger the vulnerability.

The exploit is only usable locally, and the public exploit code has been released. EPSS is not available, so the current likelihood of exploitation cannot be quantified, but the absence of a KEV listing suggests it has not yet been widely leveraged in the wild. The CVSS score indicates a moderate potential impact; attackers with local privileges could read untrusted memory or crash the process, compromising confidentiality and availability on the affected host.