Ncnn
Monthly
Out-of-bounds heap write in Tencent's ncnn neural-network inference framework allows a crafted .param model file to corrupt heap memory when parsed by Net::load_param(). Because ParamDict::load_param() and load_param_bin() validated the parsed parameter id only against the upper bound (id >= NCNN_MAX_PARAM_COUNT), a negative id indexes before the params[] array, giving an attacker a write primitive below the allocated buffer. No public exploit is identified at time of analysis and the issue is not in CISA KEV; an upstream source-code fix is available.
Integer overflow in Tencent ncnn's YUV420SP-to-RGB conversion API (ncnn::yuv420sp2rgb in src/mat_pixel.cpp) lets unauthenticated remote attackers who can get a target application to process crafted YUV image dimensions or strides trigger an undersized buffer allocation and subsequent out-of-bounds writes. The flaw affects ncnn builds up to commit 44b02dabdcf97499281a4692ebd65bad73dcf8b6 and requires the application to process attacker-controlled YUV420SP data (UI:P); realistic outcomes are denial of service or bounded heap corruption with Low confidentiality, integrity, and availability impact rather than reliable remote code execution. Publicly available exploit code exists, and the vendor has not responded to the upstream issue report.
Out-of-bounds reads in Tencent ncnn's MatMul layer can disclose process memory or crash inference workloads when a crafted model drives matmul_transb with inconsistent dimensions. Affected builds are those at or before commit 44b02dabdcf97499281a4692ebd65bad73dcf8b6, and exploitation requires the victim to load and run an attacker-supplied model (UI:P); no KEV entry exists, but publicly available exploit code exists. The issue is genuine but moderate severity, with impact bounded to memory disclosure and possible denial of service rather than code execution.
Out-of-bounds heap write in Tencent's ncnn neural-network inference framework allows a crafted .param model file to corrupt heap memory when parsed by Net::load_param(). Because ParamDict::load_param() and load_param_bin() validated the parsed parameter id only against the upper bound (id >= NCNN_MAX_PARAM_COUNT), a negative id indexes before the params[] array, giving an attacker a write primitive below the allocated buffer. No public exploit is identified at time of analysis and the issue is not in CISA KEV; an upstream source-code fix is available.
Integer overflow in Tencent ncnn's YUV420SP-to-RGB conversion API (ncnn::yuv420sp2rgb in src/mat_pixel.cpp) lets unauthenticated remote attackers who can get a target application to process crafted YUV image dimensions or strides trigger an undersized buffer allocation and subsequent out-of-bounds writes. The flaw affects ncnn builds up to commit 44b02dabdcf97499281a4692ebd65bad73dcf8b6 and requires the application to process attacker-controlled YUV420SP data (UI:P); realistic outcomes are denial of service or bounded heap corruption with Low confidentiality, integrity, and availability impact rather than reliable remote code execution. Publicly available exploit code exists, and the vendor has not responded to the upstream issue report.
Out-of-bounds reads in Tencent ncnn's MatMul layer can disclose process memory or crash inference workloads when a crafted model drives matmul_transb with inconsistent dimensions. Affected builds are those at or before commit 44b02dabdcf97499281a4692ebd65bad73dcf8b6, and exploitation requires the victim to load and run an attacker-supplied model (UI:P); no KEV entry exists, but publicly available exploit code exists. The issue is genuine but moderate severity, with impact bounded to memory disclosure and possible denial of service rather than code execution.