Beyond F1: Evaluating Coverage and Failure Recovery in AI Model Security Scanners

The growing reliance on pre-trained Machine Learning (ML) models has introduced new attack surfaces. Recent vulnerabilities demonstrate that malicious behavior can be embedded within model artifacts, often bypassing existing defenses. Current model-scanning solutions primarily rely on static, format-specific rules or known attack signatures, which limit their ability to generalize across frameworks and to detect novel exploitation paths. In contrast, we propose a solution that focuses on the effects an attack has on the host system executing the model and builds on foundational intuitions about ML model execution. In particular, we observe that ML models operate within well-defined lifecycle phases and that, within each phase, interactions with the host system are highly structured and predictable. We translate these intuitions into Moat, a dynamic lifecycle-aware approach for securing ML model execution, and instantiate this design in Re-Moat, our reference implementation. We evaluate Re-Moat across multiple ML frameworks using 77,974 real-world model artifacts from the Hugging Face Hub, 31 Proofs-of-Concept (PoCs) from CVEs, and 334 models from a state-of-the-art dataset, and compare it against state-of-the-art model-scanning solutions. Our results show that our approach detects all evaluated attack classes while maintaining a close-to-zero false-positive rate, validating our intuitions and motivating dynamic analysis for securing ML model execution.

Model repositories such as Hugging Face increasingly distribute machine learning artifacts serialized with Python's pickle format, exposing users to remote code execution (RCE) risks during model loading. Recent defenses, such as PickleBall, rely on per-library policy synthesis that requires complex system setups and verified benign models, which limits scalability and generalization. In this work, we propose a lightweight, machine-learning-based scanner that detects malicious Pickle-based files without policy generation or code instrumentation. Our approach statically extracts structural and semantic features from Pickle bytecode and applies supervised and unsupervised models to classify files as benign or malicious. We construct and release a labeled dataset of 727 Pickle-based files from Hugging Face and evaluate our models on four datasets: our own, PickleBall (out-of-distribution), Hide-and-Seek (9 advanced evasive malicious models), and synthetic joblib files. Our method achieves 90.01% F1-score compared with 7.23%-62.75% achieved by the SOTA scanners (Modelscan, Fickling, ClamAV, VirusTotal) on our dataset. Furthermore, on the PickleBall data (OOD), it achieves 81.22% F1-score compared with 76.09% achieved by the PickleBall method, while remaining fully library-agnostic. Finally, we show that our method is the only one to correctly parse and classify 9/9 evasive Hide-and-Seek malicious models specially crafted to evade scanners. This demonstrates that data-driven detection can effectively and generically mitigate Pickle-based model file attacks.

Agentic software connects language models to tools that modify files and interact with external services. Developers use model-based security checks to screen external content and user requests before agents act. System One models answer developer-defined questions with probabilities over predefined answers such as safe or unsafe, but classification accuracy alone does not establish whether these probabilities support reliable automation. We evaluate Jev, Laya, Decider, and Bespoke Nimble against specialized classifiers and language-model judges across prompt-injection detection, interaction-risk judgment, and harmful-request screening, examining decision accuracy, calibration, and selective automation. (1) High overall accuracy and low average calibration error can conceal attacks classified as safe with high confidence within particular groups. Developers should test candidate models on the intended security task and examine errors within relevant attack groups. (2) Policies selected under strict miss limits allow few test inputs automatically, and choosing allow and block thresholds separately increases automation mainly through additional blocks. Developers should verify limits on missed unsafe inputs and blocked benign inputs on independent data, and report the allowed and blocked fractions separately. (3) Comparing predictions on the same inputs shows that a language-model judge can detect unsafe inputs missed by a System One model, but can also repeat the System One model's confident mistakes and falsely flag benign inputs. Developers should test which missed unsafe inputs their review rule forwards, then measure the review model's misses and benign false alarms on those inputs.