TFLite Micro OfflineMemoryAllocation OOB PoC (controlled)
Security research only. Harmless demonstration for a Huntr MFV report against TensorFlow Lite Micro.
What these files show
GetOfflinePlannedOffsets accepts OfflineMemoryAllocation metadata without validating buffer index / payload length. During MicroInterpreter::AllocateTensors():
| File | Effect |
|---|---|
offline_short_offsets.tflite |
Metadata buffer is 12 bytes (header only, nbr_tensors=1) with no offset words → heap-buffer-overflow READ in InitializeAllocationInfo under ASan when the buffer is at the end of the allocation |
offline_bad_index.tflite |
Metadata points to buffer index 99 with only 2 buffers → FlatBuffers Vector::Get assertion / crash |
No remote callbacks, no persistence, no ACE payload.
Access
Repository is gated. protectai-bot was granted access before these files were uploaded.
Related code
tensorflow/lite/micro/micro_allocation_info.cc—GetOfflinePlannedOffsets/InitializeAllocationInfo- Upstream: https://github.com/tensorflow/tflite-micro @
096563546742ba81adb6f012ab718d196a48e02d
- Downloads last month
- -
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support