modelscan-keras-bypass-poc

ModelScan Keras Scanner Bypass PoC

Vulnerability

ModelScan v0.8.8's Keras scanner (KerasLambdaDetectScan) only checks for Lambda layers at the top level of model_config["config"]["layers"]. Lambda layers nested inside sub-models (Functional, Sequential) or wrapper layers (TimeDistributed, Bidirectional) are not detected.

Bypass Techniques

1. Nested Model (nested_lambda_bypass.keras)

Lambda layer hidden inside a nested Functional sub-model. ModelScan sees the outer layer as class_name: "Functional" and skips it.

2. Wrapper Layer (wrapper_lambda_bypass.keras)

Lambda layer wrapped in a TimeDistributed wrapper. ModelScan sees class_name: "TimeDistributed" and skips it.

Root Cause

In modelscan/scanners/keras/scan.py lines 119-131:

lambda_layers = [
    layer.get("config", {}).get("function", {})
    for layer in model_config_data.get("config", {}).get("layers", {})
    if layer.get("class_name", {}) == "Lambda"
]

This only iterates the top-level layers list. It does NOT recursively check:

• Nested model configs (layers inside config.layers[n].config.layers)
• Wrapper layer configs (the layer field in TimeDistributed/Bidirectional)

Impact

When a user loads a .keras file that passes ModelScan with 0 issues:

import keras
model = keras.models.load_model("nested_lambda_bypass.keras", safe_mode=False)
# Arbitrary code executes via marshal.loads in Lambda.from_config()

The Lambda function is deserialized via:

1. Functional.from_config() → process_layer() → recursive deserialize_keras_object()
2. Lambda.from_config() → python_utils.func_load()
3. func_load() → codecs.decode(base64) → marshal.loads() → types.FunctionType() → ACE

Reproduction

pip install modelscan[keras]
modelscan -p nested_lambda_bypass.keras
# Output: No issues found (0 Lambda layers detected)
# But the file contains a malicious Lambda layer with marshal.loads code execution