Executive brief
A security vulnerability exists in whisper.cpp, a high-performance C++ implementation of OpenAI's Whisper speech recognition model. By providing a specially crafted model file, a local attacker can cause the application to crash. This results in a denial-of-service, potentially disrupting automated transcription services or applications relying on this library for audio processing.
Technical details
A vulnerability exists in the `whisper_model_load` function within `ggml/src/ggml.c` (and `src/whisper.cpp`) of whisper.cpp. The root cause is a failure to validate hyperparameters, such as `n_audio_state`, when reading from a model file. If these values are set to zero, the loader creates zero-dimension tensors with null data pointers. Subsequent operations, such as `memcpy` or `ggml_mul_mat`, trigger a null pointer dereference or a `GGML_ASSERT` failure, leading to a process crash (SIGABRT). This attack requires the ability to provide a local model file to the application. As of the advisory date, the project has been notified but a formal patch has not been confirmed.
Affected products
- ggml-org whisper.cpp up to 1.8.2
Timeline
- 2026-05-12: disclosed: Issue reported to the project maintainers via GitHub issue #3807
- 2026-06-01: advisory: CVE-2026-10298 published by VulDB/NVD