Operational Runtime Behavior Mining for Open-Source Supply Chain Security Conference: The ACM International Conference on the Foundations of Software Engineering; July 5–9, 2026; Montréal, QC, Canada Companion Proceedings of the 34th ACM Symposium on the Foundations of Software Engineering (FSE ’26)

Zhuoran Tan, Ke Xiao, Jeremy Singer, Christos Anagnostopoulos · 2026

Read Operational Runtime Behavior Mining for Open-Source Supply Chain Security Conference: The ACM International Conference on the Foundations of Software Engineering; July 5–9, 2026; Montréal, QC, Canada Companion Proceedings of the 34th ACM Symposium on the Foundations of Software Engineering (FSE ’26) by Zhuoran Tan, Ke Xiao, Jeremy Singer, Christos Anagnostopoulos (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-06948.txt.

Introduction

Zhuoran Tan and colleagues wrote a 2026 arXiv preprint. The work covers operational runtime behavior mining for Open-Source supply chain security. ArXiv posted the preprint in 2026 under its PAPERS collection. Open-source software (OSS) is a critical component of modern software systems.

Yet supply chain security remains challenging in practice due to unavailable or obfuscated source code. Consequently, security teams often rely on runtime observations collected from sandboxed executions to investigate suspicious third-party components. The work gives specialists a compact account of the problem and one proposed way to handle it.

Catalog record

CREATORZhuoran Tan, Ke Xiao, Jeremy Singer, Christos Anagnostopoulos
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE52 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2601-06948

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Operational Runtime Behavior Mining for Open-Source Supply Chain Security Conference: The ACM International Conference on the Foundations of Software Engineering; July 5–9, 2026; Montréal, QC, Canada Companion Proceedings of the 34th ACM Symposium on the Foundations of Software Engineering (FSE ’26), July 5–9, 2026, Montreal, Canada CCS: Computing methodologies Graph learning CCS: Information systems Knowledge graphs CCS: Computing methodologies Heterogeneous learning CCS: Security and privacy Supply chain threats Zhuoran Tan, Ke Xiao, Jeremy Singer, Christos Anagnostopoulos Abstract . Open-source software (OSS) is a critical component of modern software systems, yet supply chain security remains challenging in practice due to unavailable or obfuscated source code. Consequently, security teams often rely on runtime observations collected from sandboxed executions to investigate suspicious third-party components. We present HeteroGAT-Rank, an industry-oriented runtime behavior mining system that supports analyst-in-the-loop supply chain threat investigation. The system models execution-time behaviors of OSS packages as lightweight heterogeneous graphs and applies attention-based graph learning to rank behavioral patterns that are most relevant for security analysis. Rather than aiming for fully automated detection, HeteroGAT-Rank surfaces actionable runtime signals—such as file, network, and command activities—to guide manual investigation and threat hunting. To operate at ecosystem scale, the system decouples offline behavior mining from online analysis and integrates parallel graph construction for efficient processing across multiple ecosystems. An evaluation on a large-scale OSS execution dataset shows that HeteroGAT-Rank effectively highlights meaningful and interpretable behavioral indicators aligned with real-world vulnerability and attack trends, supporting practical security workflows under realistic operational constraints. 1. Introduction Open-source software (OSS) underpins modern systems, making software supply chain security a growing operational concern. Graph-based representations such as ASTs, CFGs, and dependence graphs are effective for vulnerability analysis when source code is available ( Jia et al. 2025 ; Zhang et al. 2025b ; Lu et al. 2024 ; Rozi et al. 2024 ) . In practice, however, code is often unavailable, incomplete, or obfuscated, limiting the applicability of static analysis ( Liang et al. 2024 ) . Moreover, many supply chain attacks only manifest through runtime behaviors (e.g., multi-stage execution and environment-dependent logic) that are difficult to infer from code alone ( Tartler et al. 2014 ) . As a result, security teams increasingly rely on dynamic evidence from execution monitoring, especially in closed-source and confidential environments where runtime traces may be the only viable signal ( Barr-Smith et al. 2022 ) . A key operational challenge is scale: execution tracing produces large volumes of noisy events, and analysts must quickly identify which runtime activities are most relevant for threat hunting. We therefore focus on enabling an operational understanding of runtime behavior from traces alone: rather than reconstructing program logic, we aim to prioritize suspicious behaviors such as unexpected file access, network communication, and command execution in a form that supports analyst-in-the-loop triage. We study this setting using the OSPTrack dataset ( Tan et al. 2025 ) , which contains labeled runtime execution features collected from sandboxed runs of open-source packages across five ecosystems. We represent execution traces as lightweight heterogeneous graphs, where each software instance is modeled as a star-shaped subgraph rooted at a package node and connected to runtime entities such as file paths, sockets, DNS hosts, and executed commands. This one-hop design captures direct interactions between a package and its environment while balancing expressiveness, scalability, and interpretability under industrial constraints. To analyze these graphs, we propose HeteroGAT-Rank , an attention-based heterogeneous graph learning framework built on type-aware GATv2 attention ( Brody et al. 2022 ) . HeteroGAT-Rank learns to rank runtime entities and edges by relevance to supply chain threat analysis, producing instance-specific pivots that analysts can validate in sandbox logs and translate into hunting queries. In summary, this paper makes the following contributions 1 1 1 The HeteroGAT-Rank code is available at ( ddg [n. d.] ) to facilitate reproducibility and further research. : • We present an industry-oriented framework for mining actionable runtime behavioral indicators from sandboxed executions of open-source software to support supply chain security analysis. • We propose a scalable pipeline that transforms dynamic execution data into heterogeneous graph representations across multiple ecosystems, enabling cross-ecosystem behavioral analysis. • We develop HeteroGAT-Rank , an attention-based graph learning system that prioritizes runtime behaviors for analyst-in-the-loop investigation under realistic operational constraints. • We provide an empirical evaluation and case studies demonstrating how ranked behavioral indicators can support practical threat hunting and security workflows. 2. Literature Review This work builds on prior research in software security and supply chain analysis that leverages structured representations and learning-based techniques to extract security-relevant signals. Under realistic operational constraints, related efforts broadly fall into (i) feature representation and prioritization, and (ii) supply chain security techniques based on static or dynamic analysis. 2.1. Feature Representation and Prioritization Learning-based representation has been explored to support security analysis. Yin et al. ( Yin et al. 2023 ) proposed a modality-aware graph convolutional framework to embed heterogeneous vulnerability attributes and mitigate sparsity via transfer learning, demonstrating the value of structured representations but primarily targeting offline vulnerability analysis. Luo et al. ( Luo et al. 2023 ) studied online feature selection in high-velocity streams to improve efficiency; however, such methods typically operate on flat feature spaces and offer limited support for reasoning over structured runtime behaviors. Differential analysis has also been used to reveal localized malicious modifications: Froh et al. ( Froh et al. 2023 ) and Barr et al. ( Barr-Smith et al. 2022 ) show that fine-grained differences in package updates or binaries can expose malicious changes, yet these approaches rely on paired versions/baselines and do not directly address ecosystem-scale aggregation of runtime behaviors. In contrast, our work treats learning-based ranking as an operational aid: we surface instance-specific runtime pivots for investigation, rather than optimizing predictive performance in isolation. 2.2. Software Supply Chain Security Many supply chain defenses rely on static code analysis or controlled dynamic testing. Zheng et al. ( Zheng et al. 2024 ) proposed OSCAR, a dynamic code-poisoning detection pipeline for NPM and PyPI; while effective, its per-sample simulation and fuzzing can limit scalability in large ecosystems. Other approaches are source-code-centric: Zhang et al. ( Zhang et al. 2025b ) use property graphs with BERT-based embeddings for function-level vulnerability detection, and Zhang et al. ( Zhang et al. 2025a ) combine program slicing with GRU-based models and vulnerability dictionaries to capture recurring patterns. These techniques perform well when source code is available but are less suitable for obfuscated or closed-source components and cannot capture behaviors that only manifest during execution. In contrast, we focus on sandboxed runtime behavior as an operational data source and adopt lightweight heterogeneous graphs to support analyst-in-the-loop investigation across ecosystems; HeteroGAT-Rank complements existing detection tools by prioritizing actionable runtime signals for threat hunting and manual analysis. 3. Methodology Our framework has two components (Figures 1 and 3 ): (1) we construct a per-package heterogeneous behavior graph from sandbox runtime traces; and (2) we train an attention-based GNN on a supervised subgraph classification task and reuse node/edge attentions for differential feature ranking. Per-package heterogeneous subgraph. For each package, we build a directed heterogeneous subgraph G_{r}=(V_{r},E_{r}) rooted at the package node v_{\text{pkg}} . Nodes and edges are typed (e.g., Path , DNS , CMD , Socket and their

Read the complete text