A flaw in the OptGraph::InitQuantization routine of the CUDA Optimizer component of OpenPPL ppl.nn allows an attacker to supply crafted per_channel, bit_width, tensor_max, tensor_min, and scale values. Manipulating these arguments triggers a read outside the bounds of the internal buffer, potentially exposing arbitrary memory contents. The vulnerability is classified as an out-of-bounds read (CWE‑119, CWE‑125) and, although no code execution is guaranteed, the leakage of sensitive data can compromise confidentiality and impact downstream processing components.

OpenPPL’s ppl.nn library, specifically the CUDA Optimizer submodule, is affected. The project does not use explicit versioning and specifically states that any release may be impacted, so all known instances of ppl.nn are potentially vulnerable.

The vulnerability has a CVSS score of 5.3, indicating moderate severity. EPSS data are not currently available and the issue is not listed in the CISA KEV catalog, so the probability of exploitation is uncertain but a remote attack vector is documented. Attackers can trigger the read by sending malformed quantization parameters, but no additional conditions are described. Until an official fix is released, the risk remains moderate with potential for data compromise through remote exploitation.