Summary
The Card class of skops, used for model documentation and sharing, allows arbitrary code execution. When a file other than .zip is provided to the Card class during instantiation, the internally invoked Card.get_model method silently falls back to joblib without warning. Unlike the .skops zip-based format, joblib permits unrestricted code execution, hence bypassing the security measures of skops and enabling the execution of malicious code.
Details
The Card class supports loading the model linked to the card using the get_model method. When a .skops model is provided, it uses the load function from skops, which includes security mechanisms. The Card class also supports consistent management of the trusted list, which can be passed during instance creation. As expected, if a .skops model is provided without a trusted list and an untrusted type is encountered during loading, an error is raised. This behavior is consistent with the security principles of skops.
The problem arises when a file format other than .zip is provided. As shown in the code snippet below, in this case, the joblib library is used to load the model. This happens silently, without any warning or indication that joblib is being used. This is a significant security risk because joblib does not enforce the same security measures as skops, allowing arbitrary code execution.
# from `card/_model_card.py:354-358`
try:
if zipfile.is_zipfile(model_path):
model = load(model_path, trusted=trusted)
else:
model = joblib.load(model_path)To increase the concern, get_model is actually called internally by skops during card creation, so the user does not need to call it explicitly—only to create the Card object passing a joblib file.
PoC
Consider the following example:
from skops.card import Card
card = Card("model.skops")An attacker could share a model.skops file that, despite its name, is not a .zip file. In this case, the joblib.load function is called, allowing arbitrary code execution if the file is actually a pickle-like object. This is difficult for the user to detect, as the check is based on the file’s format, not its extension or name.
This vulnerability exists regardless of the trusted list provided (or omitted) during Card instance creation, and is unaffected by any other parameters. Moreover, it occurs at the time of Card instantiation.
Attack Scenario
An attacker can craft a malicious model file that, when used to instantiate a Card object, enables arbitrary code on the victim’s machine. This requires no user interaction beyond instantiating the Card object (not even explicit loading). Given that skops is often used in collaborative environments and is designed with security in mind, this vulnerability poses a significant threat.
Attachments
The complete PoC is available on GitHub at io-no/CVE-2025-54886.
References
io-no Reporter
Summary
The
Cardclass ofskops, used for model documentation and sharing, allows arbitrary code execution. When a file other than.zipis provided to theCardclass during instantiation, the internally invokedCard.get_modelmethod silently falls back tojoblibwithout warning. Unlike the.skopszip-based format,joblibpermits unrestricted code execution, hence bypassing the security measures ofskopsand enabling the execution of malicious code.Details
The
Cardclass supports loading the model linked to the card using theget_modelmethod. When a.skopsmodel is provided, it uses theloadfunction fromskops, which includes security mechanisms. TheCardclass also supports consistent management of thetrustedlist, which can be passed during instance creation. As expected, if a.skopsmodel is provided without atrustedlist and an untrusted type is encountered during loading, an error is raised. This behavior is consistent with the security principles ofskops.The problem arises when a file format other than
.zipis provided. As shown in the code snippet below, in this case, thejobliblibrary is used to load the model. This happens silently, without any warning or indication thatjoblibis being used. This is a significant security risk becausejoblibdoes not enforce the same security measures asskops, allowing arbitrary code execution.To increase the concern,
get_modelis actually called internally byskopsduring card creation, so the user does not need to call it explicitly—only to create theCardobject passing ajoblibfile.PoC
Consider the following example:
An attacker could share a
model.skopsfile that, despite its name, is not a.zipfile. In this case, thejoblib.loadfunction is called, allowing arbitrary code execution if the file is actually a pickle-like object. This is difficult for the user to detect, as the check is based on the file’s format, not its extension or name.This vulnerability exists regardless of the
trustedlist provided (or omitted) duringCardinstance creation, and is unaffected by any other parameters. Moreover, it occurs at the time ofCardinstantiation.Attack Scenario
An attacker can craft a malicious model file that, when used to instantiate a
Cardobject, enables arbitrary code on the victim’s machine. This requires no user interaction beyond instantiating theCardobject (not even explicit loading). Given thatskopsis often used in collaborative environments and is designed with security in mind, this vulnerability poses a significant threat.Attachments
The complete PoC is available on GitHub at io-no/CVE-2025-54886.
References