Tencent ncnn EUVD-2026-104056

Exploitation requires the target application to invoke ncnn's YUV-to-RGB Conversion API (the ncnn::yuv420sp2rgb function in src/mat_pixel.cpp) on attacker-controlled image data, supplying crafted width/height/stride values that overflow an integer during buffer-size calculation. Concrete limiting factors: the victim application must actually accept and convert external YUV420SP input (UI:P / application-mediated processing), the deployment must use ncnn at or before commit 44b02dabdcf97499281a4692ebd65bad73dcf8b6, and the resulting impact is limited (Low confidentiality/integrity/availability, no scope change) - making denial of service or bounded heap corruption the realistic outcome rather than reliable remote code execution.

This is a medium-severity issue (CVSS 4.0 base 5.3) whose signals are internally consistent but point to limited real-world urgency. The vector AV:N/AC:L/PR:N/UI:P describes a remotely-reachable, unauthenticated, low-complexity flaw that nonetheless requires passive user interaction (UI:P) - the victim application must actually ingest and convert an attacker-supplied YUV420SP image via ncnn's Conversion API. Impact is uniformly Low across confidentiality, integrity and availability (VC:L/VI:L/VA:L) with no scope change, which is typical of an integer overflow (CWE-190) that may corrupt a heap buffer rather than yielding reliable code execution. The CVSS 4.0 Exploit Maturity is E:P (Proof-of-Concept), and the intelligence confirms publicly available exploit code via the GitHub issue, so 'publicly available exploit code exists' - but there is NO CISA KEV listing and NO EPSS score provided in the input, so widespread or active exploitation cannot be asserted. A noteworthy conflicting/weakening signal: this is reported solely by VulDB with the vendor (Tencent) having 'not responded yet,' meaning affected-version boundaries and impact severity are not vendor-confirmed. Net assessment: a legitimate but low-priority library bug - real because a POC exists and it is network-reachable, but de-prioritized by the required user/application interaction, the uniformly Low impact, the absence of KEV/EPSS corroboration, and its dependency on how downstream applications invoke the vulnerable conversion routine.

An attacker crafts a malicious YUV420SP image with dimensions or stride values engineered to overflow an integer during size computation in ncnn::yuv420sp2rgb, then delivers it to an application (e.g., a mobile or server-side ML pipeline) that uses ncnn to convert camera/image input to RGB. When the victim application processes the image, the integer overflow produces an undersized allocation and subsequent out-of-bounds write, leading to a crash or limited heap corruption. A public proof-of-concept is already available via the Tencent/ncnn GitHub issue (#7035).