Research area · programme
AI Governance & Security
ISO/IEC 42001 and 5259 controls, EU AI Act obligations, privacy firewalls, sandboxed agent tools. 85 papers in the vault, 32 drawn here.
32 papers drawn here · 64 curated relations touch the area · 2 authored papers. The graph opens straight on this view; the map and every other area stay one click away.
Loading the full graph…
Papers drawn in AI Governance & Security (32)
Every paper drawn in this area, sorted by curated degree. Click a title to open its ego graph on the canvas above, or follow the DOI / arXiv link to the publisher record.
| Paper | Authors | Year | Identifier / Venue | Status | Relations |
|---|---|---|---|---|---|
| Defense-in-Depth for Forensic MCP Agents: DuckDB AST Sandboxing and Context-Aware Runtime Egress Redaction Combines DuckDB SQL AST sandboxing with context-aware runtime egress redaction so diagnostic MCP agents cannot execute mutating queries or exfiltrate customer secrets. | Vidigal | 2026 | Yugabyte Applied AI & Systems Research | vault draft | 1 |
| Systems Language Trade-Offs in Production Telemetry Collectors: Controlled Evaluation of Go and Rust Across Regex Redaction, Long-Distance Zstandard, and Parquet Generation Vidigal (2026) evaluates Go vs. Rust across regex redaction, long-distance Zstandard compression, and columnar Parquet generation for YugabyteDB telemetry collectors (`yb-collector`). | Vidigal | 2026 | Yugabyte Applied AI & Systems Research | vault draft | 1 |
| Closing the AI accountability gap Proposes SMACTR, an end-to-end internal algorithmic audit framework (Scoping, Mapping, Artifact collection, Testing, Reflection) producing a documented audit trail with sign-off. | Raji et al. | 2020 | DOI:10.1145/3351095.3372873 | peer-reviewed (vault claim) | 7 |
| ISO/IEC 42001:2023 Information technology - Artificial intelligence - Management system Auditor's reading for AI-derived operational metrics: clause 9.1 requires determining what is measured, methods that ensure valid results, when, and who analyses them, with retained documented information; 7.5.3 requires control of documented information (identification, versioning, retention, access, change). | ISO/IEC | 2023 | International Organization for Standardization | identity-only record | 7 |
| Model Cards for Model Reporting Model cards: short structured documents stating intended use, out-of-scope uses, evaluation data, metrics disaggregated by relevant factors, and caveats. Under ISO 42001 a model/system card per AI system is natural evidence for A.6.2.7 technical documentation, A.9.4 intended use and A.8.2 user information. | Mitchell et al. | 2019 | DOI:10.1145/3287560.3287596 | peer-reviewed (vault claim) | 6 |
| Auditing large language models: a three-layered approach Argues LLM audits need three coordinated layers: governance audits of the provider (management system), model audits (pre-release capability/limitation evaluation), and application audits (use-case-specific, continuous). Key limitation the authors flag: no layer alone suffices and application-level risks are context-specific. | Mökander et al. | 2023 | DOI:10.1007/s43681-023-00289-2 | peer-reviewed (vault claim) | 5 |
| ISO/IEC 5259-1:2024 through 5259-5:2025 Artificial intelligence - Data quality for analytics and machine learning (ML) (Parts 1-5: Overview, Data Quality Measures, Management Requirements, Process Framework, and Data Quality Governance Framework) ISO/IEC 5259 series extends ISO/IEC 25012-style DQ characteristics to analytics/ML data, adding measures and lifecycle/governance processes (provenance, representativeness, label quality). | ISO/IEC | 2024 | International Organization for Standardization | identity-only record | 5 |
| k-ANONYMITY: A MODEL FOR PROTECTING PRIVACY Defines k-anonymity for microdata release: every quasi-identifier equivalence class must contain at least k indistinguishable records to prevent record linkage re-identification. | SWEENEY | 2002 | DOI:10.1142/S0218488502001648 | peer-reviewed (vault claim) | 5 |
| Prefix-preserving IP address anonymization: measurement-based security evaluation and a new cryptography-based scheme Xu, Fan, Ammar, and Moon (ICNP 2002) introduce Crypto-PAn, proving that keyed cryptographic prefix-preserving pseudonymization preserves subnet/topology grouping (if two IPs share a $k$-bit prefix, their anonymized IPs share a $k$-bit prefix) while resisting frequency and known-plaintext attacks across traces. | Xu et al. | 2002 | DOI:10.1109/icnp.2002.1181415 | peer-reviewed (vault claim) | 5 |
| Revealing information while preserving privacy Proves that answering too many subset-sum queries too accurately allows reconstruction of the database. The gallery publishes ~60 exact cross-cuts of the same ~67-org population; this result is why differencing across cards is a real, not theoretical, risk for any external release. | Dinur & Nissim | 2003 | DOI:10.1145/773153.773173 | peer-reviewed (vault claim) | 5 |
| Better bug reporting with better privacy Castro, Costa, and Martin (ASPLOS 2008) formalize privacy-preserving bug reporting by computing symbolic path conditions and replacing private user data (SQL literals, payloads, file contents) with synthetic or redacted values that preserve control-flow equivalence for debugging. | Castro et al. | 2008 | DOI:10.1145/1346281.1346322 | peer-reviewed (vault claim) | 4 |
| Black-Box Access is Insufficient for Rigorous AI Audits FAccT 2024: audits limited to black-box query access are weak; rigorous audits need white-box and 'outside-the-box' access (training/deployment documentation, evaluation methodology, internal records). | Casper et al. | 2024 | DOI:10.1145/3630106.3659037 | peer-reviewed (vault claim) | 4 |
| Datasheets for Datasets Datasheets document a dataset's motivation, composition, collection process, preprocessing, uses, distribution and maintenance. In ISO 42001 terms this is the evidence for A.7.2-A.7.6 (data acquisition, quality, provenance, preparation) and A.4.3 data resources. | Gebru et al. | 2021 | DOI:10.1145/3458723 | peer-reviewed (vault claim) | 4 |
| l-Diversity: Privacy Beyond k-Anonymity Shows k-anonymity fails when a k-group shares one sensitive value (homogeneity) or the attacker has background knowledge; requires l well-represented sensitive values per equivalence class. For the fleet gallery: a cell of 5 orgs that all show the same failing posture (e.g. trust-any HBA) is k-anonymous but discloses the attribute; | Machanavajjhala et al. | 2007 | DOI:10.1145/1217299.1217302 | peer-reviewed (vault claim) | 3 |
| Protecting Privacy in Software Logs: What Should Be Anonymized? Empirical study of sensitive attributes in software logs (FSE 2025) showing that sensitive tokens (credentials, IPs, emails, session IDs, user paths) leak across diverse log severities and unexpected error/exception templates rather than only dedicated auth/audit lines. | Aghili et al. | 2025 | DOI:10.1145/3715779 | peer-reviewed (vault claim) | 3 |
| Toward Trustworthy AI Development: Mechanisms for Supporting Verifiable Claims Introduces the taxonomy of institutional mechanisms (red-teaming, third-party auditing, incident sharing, bounties), software mechanisms (audit trails, interpretability, privacy-preserving ML, reproducible pipelines), and hardware mechanisms (compute accounting, secure enclaves) needed to turn high-level AI safety/governance assertions into… | Brundage et al. | 2020 | arXiv:2004.07213 | preprint | 3 |
| DESENSITIZATION: Privacy-Aware and Attack-Preserving Crash Report Ding et al. (NDSS 2020) present DESENSITIZATION, showing that crash reports can be sanitized of sensitive user inputs while preserving the exact crashing instruction, signal (`si_signo`), register state, and call-stack fault path. | Ding et al. | 2020 | DOI:10.14722/ndss.2020.24428 | peer-reviewed (vault claim) | 2 |
| EN 18286:2026 (prEN 18286) Artificial intelligence - Quality management system for EU AI Act regulatory purposes EN 18286:2026 (developed by CEN/CENELEC JTC 21 as prEN 18286, passed formal vote July 10, 2026, and cited in the Official Journal via Commission Implementing Decision (EU) 2026/2274 on October 19, 2026) is the first harmonised European standard conferring legal presumption of conformity with Article 17 (Quality Management System) of the EU AI Act… | CEN/CENELEC | 2026 | CEN-CENELEC / Official Journal of the European Union (Decision (EU) 2026/2274) | identity-only record | 2 |
| Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act) Regulation (EU) 2024/1689 (EU AI Act) establishes binding product-safety obligations for high-risk AI systems (Chapter III, Section 2: Art. 9 Risk management system, Art. 10 Data and data governance, Art. 11 Technical documentation, Art. 12 Record-keeping/logging, Art. 13 Transparency to deployers, Art. 14 Human oversight, Art. | EU Parliament & Council | 2024 | — | identity-only record | 2 |
| Sharing computer network logs for security and privacy: a motivation for new methodologies of anonymization Slagell and Yurcik (SecureComm 2005) analyze multi-field log anonymization requirements for forensic sharing, highlighting the discrepancy between rule-firing counts (how many regex/parser rules matched across a file) and artifact-level redaction accounting (how many files and total replacements occurred), as well as the risk of unverified partial… | Slagell & Yurcik | 2005 | DOI:10.1109/seccmw.2005.1588299 | peer-reviewed (vault claim) | 2 |
| t-Closeness: Privacy Beyond k-Anonymity and l-Diversity | Li et al. | 2007 | DOI:10.1109/ICDE.2007.367856 | peer-reviewed (vault claim) | 2 |
| The Algorithmic Foundations of Differential Privacy Reference text for (eps,delta)-DP, Laplace/Gaussian mechanisms, composition and sparse vector. For fleet benchmarks: org-level (user-level) sensitivity bounding plus Laplace noise on counts, with a fixed privacy budget per release, is the standard way to publish small-cell rates externally; composition across 70 cards is the practical constraint. | Dwork & Roth | 2014 | DOI:10.1561/0400000042 | peer-reviewed (vault claim) | 2 |
| The role mining problem | Vaidya et al. | 2007 | DOI:10.1145/1266840.1266870 | peer-reviewed (vault claim) | 2 |
| The U.S. Census Bureau Adopts Differential Privacy Census 2020 adopted DP after reconstruction experiments showed exact published tables re-identify respondents; it trades accuracy at small geographies for a formal budget. Lesson for fleet advisories: many exact cross-tabulations from one population compose into a reconstruction risk even when each table looks harmless. | Abowd | 2018 | DOI:10.1145/3219819.3226070 | peer-reviewed (vault claim) | 2 |
| AI auditing: The Broken Bus on the Road to AI Accountability Most AI audit work stops at evaluation and never reaches accountability outcomes. In ISO 42001 terms, a red tile that never opens a corrective action (9.1 without 10.2) is the broken bus. | Birhane et al. | 2024 | arXiv:2401.14462 | preprint | 1 |
| Certificate Transparency CT is the canonical template for ybc's R16 --anchor design: an evidence producer (CA/bundle builder) must not be the sole authority over its own log; an external append-only chain (Merkle tree/SHA-256 manifest chain) makes silent tampering detectable. | Laurie | 2014 | DOI:10.1145/2668152.2668154 | peer-reviewed (vault claim) | 1 |
| Deep packet anonymization Classic EuroSec '09 work on protocol-parser-driven anonymization: instead of running expensive unanchored regexes over raw payloads, the parser decomposes records into named protocol fields and applies field-specific anonymization policies. | Foukarakis et al. | 2009 | DOI:10.1145/1519144.1519147 | peer-reviewed (vault claim) | 1 |
| GPTFUZZER: Red Teaming Large Language Models with Auto-Generated Jailbreak Prompts GPTFUZZER: automated red-team prompt fuzzing; method ancestor of ybc chaos-sheet adversarial corpus | Yu et al. | 2023 | arXiv:2309.10253 | preprint | 1 |
| How Bad Can It Git? Characterizing Secret Leakage in Public GitHub Repositories Large-scale scan of public GitHub found over 100k leaked secrets; regex plus entropy plus context filters were needed. | Meli et al. | 2019 | DOI:10.14722/ndss.2019.23418 | peer-reviewed (vault claim) | 1 |
| Mining least privilege attribute based access control policies | Sanders & Yue | 2019 | DOI:10.1145/3359789.3359805 | peer-reviewed (vault claim) | 1 |
| PII-TRACE: A Benchmark for Context-Aware PII Detection in Multi-Turn LLM Conversations Introduces PII-TRACE, a multi-turn conversational benchmark evaluating context-dependent personally identifiable information and secret leakage across agent interactions. | Zhang et al. | 2026 | arXiv:2609.22200 | preprint | 1 |
| Psychological Safety and Learning Behavior in Work Teams Team psychological safety predicts learning behaviour (asking for help, discussing errors), which mediates team performance; coaching-oriented leadership builds it. A blameless review list and team-level aggregates support it; name-bearing load charts visible to peers can undermine it. | Edmondson | 1999 | DOI:10.2307/2666999 | peer-reviewed (vault claim) | 1 |
Curated semantic relation notes touching Governance (44)
complements ↔ Complacency and Bias in Human Use of Automation: An Attentional Integration ↔ Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act)
Provides the human-factors empirical foundation for EU AI Act Article 14(4)(b) automation bias mitigation and ISO/IEC 42001 Annex A.9 responsible use.
Drafted by: agent draft (gemini)
complements ↔ Data Validation for Machine Learning ↔ ISO/IEC 5259-1:2024 through 5259-5:2025 Artificial intelligence - Data quality for analytics and machine learning (ML) (Parts 1-5: Overview, Data Quality Measures, Management Requirements, Process Framework, and Data Quality Governance Framework)
Implements automated batch schema validation, drift bounds, and training-serving skew checks that operationalize ISO/IEC 5259-2/4 and ISO/IEC 42001 A.7.4/A.7.6.
Drafted by: agent draft (gemini)
complements ↔ EN 18286:2026 (prEN 18286) Artificial intelligence - Quality management system for EU AI Act regulatory purposes ↔ ISO/IEC 42001:2023 Information technology - Artificial intelligence - Management system
EN 18286:2026 layers product-centric EU AI Act Article 17 QMS conformity (Annex ZA -> Arts. 11, 17, 72) onto organization-centric ISO/IEC 42001 (mapped in Annex D).
Drafted by: agent draft (gemini)
operationalizes → EN 18286:2026 (prEN 18286) Artificial intelligence - Quality management system for EU AI Act regulatory purposes → Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act)
EN 18286:2026 (Commission Implementing Decision (EU) 2026/2274) is the harmonised standard providing presumption of conformity for EU AI Act Article 17.
Drafted by: agent draft (gemini)
complements ↔ Lack of AI-Ready Data Puts AI Projects at Risk (60% Abandonment Forecast Through 2026) & Gartner Data and Analytics Summit 2026 Findings on Governed Semantic Layers for AI Agents ↔ ISO/IEC 5259-1:2024 through 5259-5:2025 Artificial intelligence - Data quality for analytics and machine learning (ML) (Parts 1-5: Overview, Data Quality Measures, Management Requirements, Process Framework, and Data Quality Governance Framework)
Empirical 60% AI project abandonment rate through 2026 motivating governed semantic layers and ISO 5259 data readiness
Drafted by: seed curator
operationalizes → ISO/IEC 5259-1:2024 through 5259-5:2025 Artificial intelligence - Data quality for analytics and machine learning (ML) (Parts 1-5: Overview, Data Quality Measures, Management Requirements, Process Framework, and Data Quality Governance Framework) → ISO/IEC 42001:2023 Information technology - Artificial intelligence - Management system
ISO/IEC 5259-1..5:2025 operationalizes ISO/IEC 42001:2023 Annex A.7 (Data for AI systems) across quality, lineage, and governance
Drafted by: seed curator
operationalizes → ISO/IEC 5259-1:2024 through 5259-5:2025 Artificial intelligence - Data quality for analytics and machine learning (ML) (Parts 1-5: Overview, Data Quality Measures, Management Requirements, Process Framework, and Data Quality Governance Framework) → Designing data governance
ISO/IEC 5259-1..5 operationalizes Khatri & Brown's Data Quality, Metadata, and Data Lifecycle decision domains for analytics and machine learning.
Drafted by: agent draft (gemini)
extends → t-Closeness: Privacy Beyond k-Anonymity and l-Diversity → l-Diversity: Privacy Beyond k-Anonymity
Extends l-diversity by bounding the Earth Mover's Distance (t-closeness) between equivalence-class sensitive attributes and the global population distribution.
Drafted by: agent draft (gemini)
operationalizes → The U.S. Census Bureau Adopts Differential Privacy → Calibrating Noise to Sensitivity in Private Data Analysis
Operationalizes formal differential privacy budgets and top-down hierarchical noise injection across the 2020 US Decennial Census disclosure avoidance system.
Drafted by: agent draft (gemini)
extends → l-Diversity: Privacy Beyond k-Anonymity → k-ANONYMITY: A MODEL FOR PROTECTING PRIVACY
Extends k-anonymity with intra-equivalence-class l-diversity to defend against homogeneity and background-knowledge attribute disclosure attacks.
Drafted by: agent draft (gemini)
complements ↔ Model Cards for Model Reporting ↔ Datasheets for Datasets
Model cards document models; datasheets document datasets; together they cover ISO 42001 A.6.2.7 and A.7 documentation evidence.
Drafted by: agent draft (claude)
complements ↔ Auditing large language models: a three-layered approach ↔ ISO/IEC 42001:2023 Information technology - Artificial intelligence - Management system
Three-layered LLM auditing (governance, model, application) maps onto ISO/IEC 42001 A.10 (suppliers), A.6.2.4 (V&V), and A.6.2.6 (operation and monitoring).
Drafted by: agent draft (gemini)
extends → Auditing large language models: a three-layered approach → Toward Trustworthy AI Development: Mechanisms for Supporting Verifiable Claims
Extends Brundage et al.'s institutional verification mechanisms into a three-layered audit architecture covering governance, model, and application audits for large language models.
Drafted by: agent draft (gemini)
extends → Auditing large language models: a three-layered approach → Closing the AI accountability gap
Extends single-organisation internal audit to a three-layer (governance/model/application) scheme specific to LLMs.
Drafted by: agent draft (claude)
complements ↔ A Comparative Study of Software Secrets Reporting by Secret Detection Tools ↔ How Bad Can It Git? Characterizing Secret Leakage in Public GitHub Repositories
Benchmarks secret-scanning detection tools across precision and recall, complementing Meli et al.'s empirical measurement of secret leakage across public Git repositories.
Drafted by: agent draft (gemini)
extends → The devil and packet trace anonymization → Prefix-preserving IP address anonymization: measurement-based security evaluation and a new cryptography-based scheme
Applies prefix-preserving IP pseudonymization alongside multi-layer payload/header transformation and highlights operational verification traps.
Drafted by: agent draft (gemini)
complements ↔ Closing the AI accountability gap ↔ ISO/IEC 42001:2023 Information technology - Artificial intelligence - Management system
SMACTR internal audit trail complements the management-system audit of cl. 9.2 with artifact-level evidence
Drafted by: agent draft (gemini)
uses method → Closing the AI accountability gap → Model Cards for Model Reporting
SMACTR Artifact Collection stage requires model cards and datasheets as audit artifacts.
Drafted by: agent draft (claude)
extends → Mining least privilege attribute based access control policies → The role mining problem
Least-privilege policy mining from audit logs
Drafted by: agent draft (claude)
complements ↔ Ten Simple Rules for Reproducible Computational Research ↔ ISO/IEC 42001:2023 Information technology - Artificial intelligence - Management system
Reproducibility rules give testable criteria for 'methods that ensure valid results' (cl. 9.1) and controlled documented information (cl. 7.5) when notebooks are management evidence
Drafted by: agent draft (gemini)
complements ↔ Hidden Technical Debt in Machine Learning Systems ↔ ISO/IEC 42001:2023 Information technology - Artificial intelligence - Management system
CACE entanglement and unstable data dependencies explain why ISO/IEC 42001 Clause 6.3/8.2 change management and A.6.2.4 V&V must gate prompts, schemas, and data together.
Drafted by: agent draft (gemini)
contrasts with ↔ LogJack: Indirect Prompt Injection Through Cloud Logs Against LLM Debugging Agents ↔ GPTFUZZER: Red Teaming Large Language Models with Auto-Generated Jailbreak Prompts
Data-plane injection via log payloads vs prompt-plane adversarial fuzzing: two attack surfaces of LLM tool stacks
Drafted by: agent draft (gemini)
complements ↔ Sharing computer network logs for security and privacy: a motivation for new methodologies of anonymization ↔ The devil and packet trace anonymization
Establishes multi-level log anonymization accounting and motivates reconciling rule-level vs artifact-level scrubbing with fail-closed verification.
Drafted by: agent draft (gemini)
complements ↔ The role mining problem ↔ Detecting anomalous access patterns in relational databases
Role mining vs role profiling from SQL logs; both derive roles from observed access
Drafted by: agent draft (claude)
Full vault context
This page lists the 32 drawn papers and 44 semantic relation notes touching AI Governance & Security. For the site-wide four evidence chains, 3 explicit refutations, retracted claim, and all 23 authored records, return to The Research Network overview →
Take the data
The graph you are looking at is a static JSON extract of the vault, served from this origin with no runtime requests. Every count on this page is reproducible from these files and scripts/build-research-graph.mjs.