CVE-2026-47749: stable-diffusion.cpp: heap overflow in .ckpt model parser

stable-diffusion.cpp is a pure C/C++ library for running diffusion model (Stable Diffusion, Flux, Wan, Qwen Image, Z-Image, and more) inference. Versions prior to master-584-0a7ae07 are vulnerable to heap buffer overflow in SHORT_BINUNICODE parsing for PyTorch checkpoint files. The pickle .ckpt parser in src/model.cpp contained a heap buffer overflow vulnerability in the SHORT_BINUNICODE opcode handler. The issue was caused by sign confusion on the opcode length field. A crafted .ckpt file could trigger memcpy with a very large length derived from a negative signed value, causing immediate heap corruption. Any application using affected stable-diffusion.cpp releases to load untrusted .ckpt model files could be vulnerable. A malicious checkpoint file could cause heap corruption through memcpy with an attacker-controlled length. This may lead to process crash and could potentially be leveraged for code execution depending on heap layout. The attack requires the victim or application to load a .ckpt file from an untrusted source, such as a downloaded model from a model sharing site. The issue has been resolved in version master-584-0a7ae07. If developers are unable to immediately update their applications they can work around this issue by not loading .ckpt checkpoint files from untrusted sources, and referring to trusted model sources and safer formats such as .safetensors where possible.

An adversary constructs a .ckpt file containing a malformed SHORT_BINUNICODE opcode where the two-byte length field is set to a negative signed value such as -1 (0xFFFF). The vulnerable handler in stable-diffusion.cpp's src/model.cpp interprets this as a very large unsigned integer (~65535), passing it directly to memcpy as the copy length. The adversary publishes this file to a popular model sharing platform under a compelling identity — a fine-tuned character model, a style LoRA, or a checkpoint with fabricated positive reviews. A developer or operator downloads the file as part of normal model evaluation, loads it through their stable-diffusion.cpp application, and the heap buffer overflow fires before any model weights are ever read. The resulting heap corruption causes either an immediate process crash (reliable denial of service) or, with crafted heap grooming in the malicious pickle stream, arbitrary write primitives enabling code execution on the host system.