{ "schema": "lri.security-direct-paper-evidence.v1", "entry_count": 12, "entries": [ { "card_id": "molot_system_card", "title": "MOLOT System Card: Malicious Operational Logic Observation Transformer", "source_url": "https://arxiv.org/abs/2606.07792", "primary_public_url": "https://arxiv.org/abs/2606.07792", "primary_self_reference": null, "markdown_card": "sources/resource_cards/01_molot_system_card.md", "id": "molot_system_card", "evidence_level": "L2_direct_public_evidence", "direct_evidence_scope": "直接支撑 MOLOT 作为 SAST 静态恶意代码检测系统、静态调用图行为序列、BERT/Transformer 分类、SHAP 源码位置映射、Python/JavaScript 包评估和 Open Malicious-Code Bench。", "extracted_claims": [ "MOLOT targets SAST settings where package metadata, maintainer history, and dynamic traces may be unavailable or unreliable.", "MOLOT represents source code as behavior sequences derived from static call graphs.", "MOLOT includes an explanation stage ranking suspicious activities and mapping them back to source locations.", "MOLOT is evaluated on Python and JavaScript packages and compared with open-source detection tools." ], "does_not_support": [ "FCD C/C++ performance table", "PDG binary clone-detection precision/recall", "LRI Agent Gate runtime metrics before implementation" ] }, { "card_id": "cerebro_behavior_sequence", "title": "Malicious Package Detection in NPM and PyPI using a Single Model of Malicious Behavior Sequence", "source_url": "https://arxiv.org/abs/2309.02637", "primary_public_url": "https://arxiv.org/abs/2309.02637", "primary_self_reference": null, "markdown_card": "sources/resource_cards/02_cerebro_behavior_sequence.md", "id": "cerebro_behavior_sequence", "evidence_level": "L1_public_method_background", "direct_evidence_scope": "直接支撑 NPM/PyPI 恶意包可以通过高层恶意行为抽象、行为序列和 BERT 统一建模。", "extracted_claims": [ "Cerebro detects malicious packages in NPM and PyPI.", "It curates high-level malicious behavior abstractions, organizes features into behavior sequences, and fine-tunes BERT.", "It reported discovery of new malicious PyPI and NPM packages." ], "does_not_support": [ "MOLOT 的具体性能", "FCD/PDG C/C++ 表", "Agent Gate 工程表现" ] }, { "card_id": "python_malware_benchmark", "title": "A Benchmark Comparison of Python Malware Detection Approaches", "source_url": "https://arxiv.org/abs/2209.13288", "primary_public_url": "https://arxiv.org/abs/2209.13288", "primary_self_reference": null, "markdown_card": "sources/resource_cards/03_python_malware_benchmark_bandit4mal_ossgadget.md", "id": "python_malware_benchmark", "evidence_level": "L2_direct_baseline_evidence", "direct_evidence_scope": "直接支撑 PyPI 恶意包检测 benchmark、Bandit4Mal/OSSGadget/PyPI checks 对比、以及仓库 moderation 对误报率极其敏感的约束。", "extracted_claims": [ "The paper builds a benchmark for Python/PyPI malware detection.", "It compares PyPI checks, Bandit4Mal, and OSSGadget OSS Detect Backdoor.", "It reports repository administrators consider even a 0.01% false-positive rate too high in this setting." ], "does_not_support": [ "MOLOT 方法本身", "NPM benchmark 2026 结论", "PDG/FCD 表" ] }, { "card_id": "npm_malicious_detection_benchmark", "title": "Understanding NPM Malicious Package Detection: A Benchmark-Driven Empirical Analysis", "source_url": "https://arxiv.org/abs/2603.27549", "primary_public_url": "https://arxiv.org/abs/2603.27549", "primary_self_reference": null, "markdown_card": "sources/resource_cards/04_npm_malicious_detection_benchmark.md", "id": "npm_malicious_detection_benchmark", "evidence_level": "L2_direct_benchmark_evidence", "direct_evidence_scope": "直接支撑 NPM 恶意包 benchmark、行为类别、规避技术、多工具对比和行为链优先判断。", "extracted_claims": [ "The paper builds a dataset with malicious and benign NPM packages.", "It annotates behavior categories and evasion techniques.", "It evaluates multiple tools and variants.", "It argues behavioral chains disambiguate malicious purpose better than isolated API calls." ], "does_not_support": [ "Python/PyPI 性能", "MOLOT 论文指标", "FCD/PDG 表" ] }, { "card_id": "oss_malicious_package_analysis", "title": "OSS Malicious Package Analysis in the Wild", "source_url": "https://arxiv.org/abs/2404.04991", "primary_public_url": "https://arxiv.org/abs/2404.04991", "primary_self_reference": null, "markdown_card": "sources/resource_cards/05_oss_malicious_package_analysis.md", "id": "oss_malicious_package_analysis", "evidence_level": "L2_direct_dataset_evidence", "direct_evidence_scope": "直接支撑 OSS 恶意包大规模数据集、多来源收集、生命周期和攻击活动上下文的重要性。", "extracted_claims": [ "The paper builds a dataset of 23,425 malicious packages from scattered online sources.", "It uses a knowledge graph to represent the OSS malware corpus.", "It describes a lifecycle of changing, release, detection, and removal, and highlights package variants." ], "does_not_support": [ "某个检测器精度", "MOLOT 方法细节", "FCD/PDG 表" ] }, { "card_id": "backstabbers_knife_collection", "title": "Backstabber's Knife Collection: A Review of Open Source Software Supply Chain Attacks", "source_url": "https://arxiv.org/abs/2005.09535", "primary_public_url": "https://arxiv.org/abs/2005.09535", "primary_self_reference": null, "markdown_card": "sources/resource_cards/06_backstabbers_knife_collection.md", "id": "backstabbers_knife_collection", "evidence_level": "L1_public_supply_chain_taxonomy", "direct_evidence_scope": "直接支撑真实 OSS 供应链攻击样本、npm/PyPI/RubyGems 恶意包、攻击树和执行时机分析。", "extracted_claims": [ "The paper reviews supply-chain attacks involving injection of malicious code into software packages.", "It presents 174 malicious packages from npm, PyPI, and RubyGems.", "It provides attack trees for injection into dependency trees and execution at different times/conditions." ], "does_not_support": [ "2026 新型攻击覆盖", "MOLOT 性能", "Agent Gate 表现" ] }, { "card_id": "unveiling_malicious_logic", "title": "Unveiling Malicious Logic: Towards a Statement-Level Taxonomy and Dataset for Securing Python Packages", "source_url": "https://arxiv.org/abs/2512.12559", "primary_public_url": "https://arxiv.org/abs/2512.12559", "primary_self_reference": null, "markdown_card": "sources/resource_cards/07_unveiling_malicious_logic_statement_taxonomy.md", "id": "unveiling_malicious_logic", "evidence_level": "L2_direct_statement_level_evidence", "direct_evidence_scope": "直接支撑 Python 恶意包语句级标注、47 个恶意指标、7 类类型和 recurring attack workflow 的定位分析。", "extracted_claims": [ "The paper builds a statement-level dataset for malicious Python packages.", "It derives 47 malicious indicators across 7 types.", "It applies sequential pattern mining to recurring indicator sequences." ], "does_not_support": [ "JavaScript/NPM 覆盖", "MOLOT 性能", "PDG/FCD 表" ] }, { "card_id": "llm_pypi_malicious_detection", "title": "Detecting Malicious Source Code in PyPI Packages with LLMs: Does RAG Come in Handy?", "source_url": "https://arxiv.org/abs/2504.13769", "primary_public_url": "https://arxiv.org/abs/2504.13769", "primary_self_reference": null, "markdown_card": "sources/resource_cards/08_llm_pypi_malicious_detection.md", "id": "llm_pypi_malicious_detection", "evidence_level": "L1_public_llm_labeling_evidence", "direct_evidence_scope": "直接支撑 LLM、RAG、few-shot 在 PyPI 恶意源码检测中的比较，以及 RAG 不一定带来改进的经验结果。", "extracted_claims": [ "The paper evaluates LLMs, RAG, and few-shot learning for malicious source code detection in PyPI packages.", "It reports RAG performs only modestly in its evaluation while few-shot learning performs better.", "It integrates YARA rules, GitHub Security Advisories, and malicious code snippets." ], "does_not_support": [ "LLM 可作为自动 block 的唯一依据", "MOLOT 论文指标", "FCD/PDG 表" ] }, { "card_id": "pdg_holmes_graph_siamese", "title": "Modeling Functional Similarity in Source Code with Graph-Based Siamese Networks", "source_url": "https://arxiv.org/abs/2011.11228", "primary_public_url": "https://arxiv.org/abs/2011.11228", "primary_self_reference": null, "markdown_card": "sources/resource_cards/09_pdg_holmes_graph_siamese_clone_detection.md", "id": "pdg_holmes_graph_siamese", "evidence_level": "L1_public_graph_clone_background", "direct_evidence_scope": "直接支撑程序依赖图和图神经网络用于语义代码 clone detection。", "extracted_claims": [ "The paper addresses semantic code clone detection using program dependency graphs and geometric neural networks.", "It develops HOLMES, a graph-based Siamese prototype.", "It evaluates on clone benchmarks and compares against TBCCD." ], "does_not_support": [ "恶意包检测精度", "FCD C/C++ 表", "MOLOT 性能" ] }, { "card_id": "source_code_is_graph", "title": "Source Code is a Graph, Not a Sequence: A Cross-Lingual Perspective on Code Clone Detection", "source_url": "https://arxiv.org/abs/2312.16488", "primary_public_url": "https://arxiv.org/abs/2312.16488", "primary_self_reference": null, "markdown_card": "sources/resource_cards/10_source_code_is_graph_clone_detection.md", "id": "source_code_is_graph", "evidence_level": "L1_public_graph_clone_background", "direct_evidence_scope": "直接支撑源码图表示在 code clone detection，尤其跨语言 clone 上，相比序列模型的优势。", "extracted_claims": [ "The paper argues source code is inherently a graph rather than a sequence for clone detection.", "It compares CodeBERT and CodeGraph on clone benchmarks.", "It reports graph-based methods outperform sequence-based methods, especially cross-lingually." ], "does_not_support": [ "恶意代码检测精度", "MOLOT 评估", "FCD C/C++ 表" ] }, { "card_id": "shap_unified_approach", "title": "A Unified Approach to Interpreting Model Predictions", "source_url": "https://arxiv.org/abs/1705.07874", "primary_public_url": "https://arxiv.org/abs/1705.07874", "primary_self_reference": null, "markdown_card": "sources/resource_cards/11_shap_unified_approach.md", "id": "shap_unified_approach", "evidence_level": "L1_public_explainability_foundation", "direct_evidence_scope": "直接支撑 SHAP 作为 feature attribution / additive explanation 方法，用于解释复杂模型预测。", "extracted_claims": [ "The paper introduces SHAP as a unified framework for interpreting model predictions.", "SHAP assigns each feature an importance value for a particular prediction.", "The framework connects additive feature-importance measures with desirable theoretical properties." ], "does_not_support": [ "恶意代码定位本身", "MOLOT 性能", "单独决定 block 策略" ] }, { "card_id": "fcd_pdg_performance_unresolved", "title": "FCD / PDG C/C++ Performance Table from User JSON", "source_url": "user_provided_json", "primary_public_url": null, "primary_self_reference": "user_provided_json", "markdown_card": "sources/resource_cards/12_fcd_pdg_performance_unresolved.md", "id": "fcd_pdg_performance_unresolved", "evidence_level": "L0_user_input_only", "direct_evidence_scope": "无直接公网论文证据；只保留为用户 JSON 中的待核验 C/C++ FCD/PDG 性能表线索。", "extracted_claims": [ "User JSON includes c-ares, jasper, openssl, and rsync rows with precision, recall, RMSE, and FCD speed.", "No public original paper/report URL has been found for that exact table in this iteration." ], "does_not_support": [ "任何已核验性能结论", "MOLOT 论文指标", "PDG/FCD 方法归属" ] } ] }