The Cost of Late Governance: A Blast-Radius Model for Pre-Promotion vs. Post-Hoc Policy Enforcement in Autonomous Multi-Agent Data Consumption
Article Sidebar
Main Article Content
Abstract
The autopilot multi-agent system is gaining immense popularity as major language models are built with external tools, shared memory and data services to accomplish tasks with minimal human oversight. This architecture transforms the governance issue in that policy breakdown which propagates beyond the initial agent that takes data. The research paper proposes a conceptual Blast-Radius Governance Model to make a comparison between pre-promotion policy enforcement and post-hoc enforcement in autonomous multi-agent data consumption. It claims that the timing of governance is a risk variable as opposed to administration choice. In this research the suggested model determines the blast radius by considering the affected-agent scope, data exposure, depth propagation, action impact, governance latency and recoverability. The analysis indicates that pre-promotion controls might maintain some limits on unsafe states prior to their reachability, but post-hoc controls do so once they have been exposed to some degree. An integrated pre-promotion validation, runtime monitoring and post-hoc monitoring is proposed as a layered model. The research paper presents a conceptual foundation which can be tested but does not purport to be empirically validated.
Article Details
This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.
References
[1] M. Janssen, P. Brous, E. Estevez, L. S. Barbosa, and T. Janowski, “Data governance: Organizing data for trustworthy Artificial Intelligence,” Gov. Inf. Q., vol. 37, no. 3, p. 101493, 2020, doi: 10.1016/j.giq.2020.101493.
[2] I. D. Raji et al., “Closing the AI accountability gap: Defining an end-to-end framework for internal algorithmic auditing,” in Proceedings of the 2020 conference on fairness, accountability, and transparency, 2020, pp. 33–44. doi: https://doi.org/10.1145/3351095.3372873.
[3] D. Dell’Anna, M. Dastani, and F. Dalpiaz, “Runtime revision of sanctions in normative multi-agent systems,” Auton. Agent. Multi. Agent. Syst., vol. 34, no. 2, p. 43, 2020, doi: 10.1007/s10458-020-09465-8.
[4] J. Mökander, J. Morley, M. Taddeo, and L. Floridi, “Ethics-Based Auditing of Automated Decision-Making Systems: Nature, Scope, and Limitations,” Sci. Eng. Ethics, vol. 27, no. 4, p. 44, 2021, doi: 10.1007/s11948-021-00319-4.
[5] J. Chen et al., “Llmarena: Assessing capabilities of large language models in dynamic multi-agent environments,” in Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2024, pp. 13055–13077. doi: https://doi.org/10.18653/v1/2024.acl-long.705.
[6] Y. Wang, D. Xue, S. Zhang, and S. Qian, “Badagent: Inserting and activating backdoor attacks in llm agents,” in Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2024, pp. 9811–9827. doi: https://doi.org/10.18653/v1/2024.acl-long.530.
[7] Q. Zhan, Z. Liang, Z. Ying, and D. Kang, “Injecagent: Benchmarking indirect prompt injections in tool-integrated large language model agents,” in Findings of the Association for Computational Linguistics: ACL 2024, 2024, pp. 10471–10506. doi: https://doi.org/10.18653/v1/2024.findings-acl.624.
[8] B. W. Wirtz, J. C. Weyerer, and I. Kehl, “Governance of artificial intelligence: A risk and guideline-based integrative framework,” Gov. Inf. Q., vol. 39, no. 4, p. 101685, 2022, doi: 10.1016/j.giq.2022.101685.
[9] L. Schmitt, “Mapping global AI governance: a nascent regime in a fragmented landscape,” AI Ethics, vol. 2, no. 2, pp. 303–314, 2022, doi: 10.1007/s43681-021-00083-y.
[10] M. Fisher, V. Mascardi, K. Y. Rozier, B.-H. Schlingloff, M. Winikoff, and N. Yorke-Smith, “Towards a framework for certification of reliable autonomous systems,” Auton. Agent. Multi. Agent. Syst., vol. 35, no. 1, p. 8, 2021, doi: 10.1007/s10458-020-09487-2.
[11] L. T. van Binsbergen, M. G. Kebede, J. Baugh, T. van Engers, and D. G. van Vuurden, “Dynamic generation of access control policies from social policies,” Procedia Comput. Sci., vol. 198, pp. 140–147, 2022, doi: 10.1016/j.procs.2021.12.221.
[12] B. Otto, M. ten Hompel, and S. Wrobel, Designing Data Spaces: The Ecosystem Approach to Competitive Advantage. Springer, 2022. doi: 10.1007/978-3-030-93975-5.
[13] S. Oduro, E. Moss, and J. Metcalf, “Obligations to assess: Recent trends in AI accountability regulations,” Patterns, vol. 3, no. 11, p. 100608, 2022, doi: 10.1016/j.patter.2022.100608.
[14] J. Mökander and M. Axente, “Ethics-based auditing of automated decision-making systems: intervention points and policy implications,” AI Soc., vol. 38, no. 1, pp. 153–171, 2023, doi: 10.1007/s00146-021-01286-x.
[15] C. Novelli, M. Taddeo, and L. Floridi, “Accountability in artificial intelligence: what it is and how it works,” AI Soc., vol. 39, no. 4, pp. 1871–1882, 2024, doi: 10.1007/s00146-023-01635-y.
[16] G. Malgieri and F. Pasquale, “Licensing high-risk artificial intelligence: Toward ex ante justification for a disruptive technology,” Comput. Law Secur. Rev., vol. 52, p. 105899, 2024, doi: 10.1016/j.clsr.2023.105899.
[17] W. Hua et al., “Trustagent: Towards safe and trustworthy llm-based agents,” in Findings of the association for computational linguistics: EMNLP 2024, 2024, pp. 10000–10016. doi: https://doi.org/10.18653/v1/2024.findings-emnlp.585.
[18] T. Sadhu, A. Pesaranghader, Y. Chen, and D. H. Yi, “Athena: Safe autonomous agents with verbal contrastive learning,” in Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing: Industry Track, 2024, pp. 1121–1130. doi: https://doi.org/10.18653/v1/2024.emnlp-industry.84.
[19] N. Zwerdling, D. Boaz, E. Rabinovich, G. Uziel, D. Amid, and A. A. Tavor, “Towards enforcing company policy adherence in agentic workflows,” in Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing: Industry Track, 2025, pp. 595–606. doi: https://doi.org/10.18653/v1/2025.emnlp-industry.41.
[20] J. Ankam, “Federated Governance Without Central Bottlenecks: A Policy Propagation Model for Data Mesh,” SSRN Electron. J., 2022, doi: 10.2139/ssrn.7323259.
[21] P. Danassis, Z. D. Erden, and B. Faltings, “Exploiting environmental signals to enable policy correlation in large-scale decentralized systems,” Auton. Agent. Multi. Agent. Syst., vol. 36, no. 1, p. 13, 2022, doi: https://doi.org/10.1007/s10458-021-09541-7.
[22] R. Shahroz, Z. Tan, S. Yun, C. Fleming, and T. Chen, “Agents under siege: Breaking pragmatic multi-agent llm systems with optimized prompt attacks,” in Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2025, pp. 9661–9674. doi: https://doi.org/10.18653/v1/2025.acl-long.476.
[23] J. Ankam, “Retrieval-Ready Pipelines: Benchmarking Chunking and Embedding Strategies for Enterprise RAG,” SSRN Electron. J., 2023, doi: 10.2139/ssrn.7323278.
[24] D. Agarwal, A. R. Fabbri, B. Risher, P. Laban, S. Joty, and C.-S. Wu, “Prompt Leakage effect and mitigation strategies for multi-turn LLM Applications,” Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing: Industry Track. pp. 1255–1275, 2024. doi: 10.18653/v1/2024.emnlp-industry.94.
[25] T. Yuan et al., “R-Judge: Benchmarking Safety Risk Awareness for LLM Agents,” Findings of the Association for Computational Linguistics: EMNLP 2024. pp. 1467–1490, 2024. doi: 10.18653/v1/2024.findings-emnlp.79.
[26] J. Ankam, “ACID-Lake: A Metadata-Driven Concurrency Control Protocol for Multi-Writer Transactions on Object Storage,” SSRN Electron. J., 2021, doi: 10.2139/ssrn.7323258.
[27] N. Palladino, “A ‘biased’ emerging governance regime for artificial intelligence? How AI ethics get skewed moving from principles to practices,” Telecomm. Policy, vol. 47, no. 5, p. 102479, 2023, doi: 10.1016/j.telpol.2022.102479.