OpenVINO TensorFlow SavedModel HashTable Control-Dependency Misbinding
This package preserves a locally confirmed OpenVINO TensorFlow frontend issue where a crafted SavedModel changes inference output by moving a control dependency into the LookupTableImportV2 input list for HashTableV2.
Current status on Thursday, July 16, 2026:
- exploit confirmed locally
- affected versions confirmed:
2025.4.1and2026.2.1 - victim workflow confirmed:
ov.convert_model(<savedmodel_dir>) - no matching public Hugging Face model repo was found for this exact primitive during current prior-art search
Included models
model/clean_savedmodel/- baseline SavedModel; expected output
[222]
- baseline SavedModel; expected output
model/crafted_savedmodel/- crafted SavedModel where the raw input list is
["HashTable", "^bogus", "keys", "values"] - expected output
[5]
- crafted SavedModel where the raw input list is
model/tail_control_savedmodel/- negative control where the raw input list is
["HashTable", "keys", "values", "^bogus"] - expected output
[222]
- negative control where the raw input list is
Reproduction
From this package directory:
PYTHONPATH=/path/to/openvino-2026.2.1-site-packages python3 reproduce/run_openvino_tf_savedmodel.py model/clean_savedmodel
PYTHONPATH=/path/to/openvino-2026.2.1-site-packages python3 reproduce/run_openvino_tf_savedmodel.py model/crafted_savedmodel
PYTHONPATH=/path/to/openvino-2026.2.1-site-packages python3 reproduce/run_openvino_tf_savedmodel.py model/tail_control_savedmodel
The runner defaults to ov.convert_model(...), which is the primary victim workflow used for this PoC. The explicit TensorFlow frontend path also reproduces:
PYTHONPATH=/path/to/openvino-2026.2.1-site-packages python3 reproduce/run_openvino_tf_savedmodel.py --api frontend model/crafted_savedmodel
To regenerate the model set deterministically:
python3 generator/generate_tf_savedmodel_lookup_misbinding.py /tmp/tf_savedmodel_lookup_misbinding
Expected outputs:
- clean:
[222] - crafted:
[5] - tail control:
[222]
Equivalent high-level victim API:
import openvino as ov
model = ov.convert_model("model/crafted_savedmodel")
compiled = ov.Core().compile_model(model, "CPU")
print(next(iter(compiled({}).values())).tolist())
Root cause summary
- SavedModel pre-analysis graph builder:
- variables_index.cpp
PtrNode::associate_node()parses every rawNodeDef.input()entry, strips the^prefix, and pushes all parents into positionalinputs, including control dependencies
- SavedModel pre-analysis:
- variables_index.cpp
LookupTableImportV2checks rawnode->input_size()but then dereferences compacted positional parents viainputs[1]andinputs[2]
- Translation path:
- translate_session.cpp
- control edges are skipped from
NodeContextdata inputs
- HashTable initialization:
- hash_table.cpp
HashTableV2is preinitialized from the mismatched keys/values maps
- Inference use:
- lookup_table_find.cpp
- lookup consumes the preinitialized table and returns a silently manipulated value
Security impact
This is a model-artifact-driven silent output manipulation issue. The crafted SavedModel changes inference semantics without changing victim code or requiring non-default unsafe flags. The trigger is structural and reliable: placing a control dependency in the second LookupTableImportV2 input slot makes OpenVINO preinitialize HashTableV2 from the wrong constants, so a query that should miss and return [222] instead matches the attacker-chosen bogus key and returns [5].