{"status":"ok","message-type":"work","message-version":"1.0.0","message":{"indexed":{"date-parts":[[2025,11,3]],"date-time":"2025-11-03T04:57:31Z","timestamp":1762145851306,"version":"build-2065373602"},"reference-count":49,"publisher":"MDPI AG","issue":"8","license":[{"start":{"date-parts":[[2025,7,30]],"date-time":"2025-07-30T00:00:00Z","timestamp":1753833600000},"content-version":"vor","delay-in-days":0,"URL":"https:\/\/creativecommons.org\/licenses\/by\/4.0\/"}],"funder":[{"name":"The Future of Life Institute"},{"name":"AI Safety Camp"},{"name":"The Survival &amp; Flourishing Fund"}],"content-domain":{"domain":[],"crossmark-restriction":false},"short-container-title":["Information"],"abstract":"<jats:p>Agentic AI systems, possessing capabilities for autonomous planning and action, show great potential across diverse domains. However, their practical deployment is hindered by challenges in aligning their behavior with varied human values, complex safety requirements, and specific compliance needs. Existing alignment methodologies often falter when faced with the complex task of providing personalized context without inducing confabulation or operational inefficiencies. This paper introduces a novel solution: a \u2018superego\u2019 agent, designed as a personalized oversight mechanism for agentic AI. This system dynamically steers AI planning by referencing user-selected \u2018Creed Constitutions\u2019\u2014encapsulating diverse rule sets\u2014with adjustable adherence levels to fit non-negotiable values. A real-time compliance enforcer validates plans against these constitutions and a universal ethical floor before execution. We present a functional system, including a demonstration interface with a prototypical constitution-sharing portal, and successful integration with third-party models via the Model Context Protocol (MCP). Comprehensive benchmark evaluations (HarmBench, AgentHarm) demonstrate that our Superego agent dramatically reduces harmful outputs\u2014achieving up to a 98.3% harm score reduction and near-perfect refusal rates (e.g., 100% with Claude Sonnet 4 on AgentHarm\u2019s harmful set) for leading LLMs like Gemini 2.5 Flash and GPT-4o. This approach substantially simplifies personalized AI alignment, rendering agentic systems more reliably attuned to individual and cultural contexts, while also enabling substantial safety improvements.<\/jats:p>","DOI":"10.3390\/info16080651","type":"journal-article","created":{"date-parts":[[2025,7,30]],"date-time":"2025-07-30T14:00:41Z","timestamp":1753884041000},"page":"651","update-policy":"https:\/\/doi.org\/10.3390\/mdpi_crossmark_policy","source":"Crossref","is-referenced-by-count":1,"title":["Personalized Constitutionally-Aligned Agentic Superego: Secure AI Behavior Aligned to Diverse Human Values"],"prefix":"10.3390","volume":"16","author":[{"ORCID":"https:\/\/orcid.org\/0000-0002-4306-7577","authenticated-orcid":false,"given":"Nell","family":"Watson","sequence":"first","affiliation":[{"name":"School of Computing and Engineering, University of Gloucestershire, The Park, Cheltenham GL50 2RH, UK"}],"role":[{"role":"author","vocabulary":"crossref"}]},{"given":"Ahmed","family":"Amer","sequence":"additional","affiliation":[{"name":"Independent Researcher, Hills Rd, Cambridge CB2 8PH, UK"}],"role":[{"role":"author","vocabulary":"crossref"}]},{"given":"Evan","family":"Harris","sequence":"additional","affiliation":[{"name":"Independent Researcher, 47645 College Dr, St Mary\u2019s City, MD 20686, USA"}],"role":[{"role":"author","vocabulary":"crossref"}]},{"given":"Preeti","family":"Ravindra","sequence":"additional","affiliation":[{"name":"Independent Researcher, 4616 Henry St, Pittsburgh, PA 15213, USA"}],"role":[{"role":"author","vocabulary":"crossref"}]},{"ORCID":"https:\/\/orcid.org\/0000-0001-5699-2676","authenticated-orcid":false,"given":"Shujun","family":"Zhang","sequence":"additional","affiliation":[{"name":"School of Computing and Engineering, University of Gloucestershire, The Park, Cheltenham GL50 2RH, UK"}],"role":[{"role":"author","vocabulary":"crossref"}]}],"member":"1968","published-online":{"date-parts":[[2025,7,30]]},"reference":[{"key":"ref_1","doi-asserted-by":"crossref","first-page":"411","DOI":"10.1007\/s11023-020-09539-2","article-title":"Artificial Intelligence, Values, and Alignment","volume":"30","author":"Gabriel","year":"2020","journal-title":"Minds Mach."},{"key":"ref_2","doi-asserted-by":"crossref","first-page":"149","DOI":"10.1007\/s10676-006-0004-4","article-title":"Artificial Morality: Top-down, Bottom-up, and Hybrid Approaches","volume":"7","author":"Allen","year":"2005","journal-title":"Ethics Inf. Technol."},{"key":"ref_3","doi-asserted-by":"crossref","first-page":"185","DOI":"10.1007\/s13347-019-00354-x","article-title":"Translating Principles into Practices of Digital Ethics: Five Risks of Being Unethical","volume":"32","author":"Floridi","year":"2019","journal-title":"Philos. Technol."},{"key":"ref_4","unstructured":"Casper, S., Davies, X., Shi, C., Gilbert, T.K., Scheurer, J., Rando, J., Freedman, R., Korbak, T., Lindner, D., and Freire, P. (2023). Open Problems and Fundamental Limitations of Reinforcement Learning from Human Feedback. arXiv."},{"key":"ref_5","doi-asserted-by":"crossref","unstructured":"Wo\u017aniak, S., Koptyra, B., Janz, A., Kazienko, P., and Koco\u0144, J. (2024). Personalized Large Language Models. arXiv.","DOI":"10.1109\/ICDMW65004.2024.00071"},{"key":"ref_6","doi-asserted-by":"crossref","unstructured":"Watson, E., Viana, T., Sturgeon, B., Petersson, L., and Zhang, S. (2024). Towards an End-to-End Personal Fine-Tuning Framework for AI Value Alignment. Electronics, 13.","DOI":"10.3390\/electronics13204044"},{"key":"ref_7","unstructured":"Watson, N., and Hessami, A. (2025, July 18). Safer Agentic AI. SaferAgenticAI.org. Available online: https:\/\/www.saferagenticai.org."},{"key":"ref_8","unstructured":"Anthropic (2025, May 28). Introducing the Model Context Protocol. Available online: https:\/\/www.anthropic.com\/news\/model-context-protocol."},{"key":"ref_9","first-page":"1","article-title":"Deep Reinforcement Learning from Human Preferences","volume":"30","author":"Christiano","year":"2017","journal-title":"Adv. Neural Inf. Process. Syst."},{"key":"ref_10","doi-asserted-by":"crossref","unstructured":"Sui, P., Duede, E., Wu, S., and So, R.J. (2024). Confabulation: The Surprising Value of Large Language Model Hallucinations. arXiv.","DOI":"10.18653\/v1\/2024.acl-long.770"},{"key":"ref_11","doi-asserted-by":"crossref","unstructured":"Watson, E., Nguyen, M., Pan, S., and Zhang, S. (2025). Choice Vectors: Streamlining Personal AI Alignment Through Binary Selection. Multimodal Technol. Interact., 9.","DOI":"10.3390\/mti9030022"},{"key":"ref_12","first-page":"1","article-title":"The Ego and the Id","volume":"Volume XIX,","author":"Strachey","year":"1923","journal-title":"The Standard Edition of the Complete Psychological Works of Sigmund Freud"},{"key":"ref_13","doi-asserted-by":"crossref","unstructured":"Hosseini, E., Casto, C., Zaslavsky, N., Conwell, C., Richardson, M., and Fedorenko, E. (2024). Universality of Representation in Biological and Artificial Neural Networks. bioRxiv.","DOI":"10.1101\/2024.12.26.629294"},{"key":"ref_14","doi-asserted-by":"crossref","first-page":"389","DOI":"10.1016\/j.neuron.2004.09.027","article-title":"The neural bases of cognitive conflict and control in moral judgment","volume":"44","author":"Greene","year":"2004","journal-title":"Neuron"},{"key":"ref_15","doi-asserted-by":"crossref","first-page":"276","DOI":"10.1093\/cercor\/bhn080","article-title":"The neural basis of human social values: Evidence from functional MRI","volume":"19","author":"Zahn","year":"2009","journal-title":"Cereb. Cortex"},{"key":"ref_16","unstructured":"Baars, B.J. (1988). A Cognitive Theory of Consciousness: The Workspace of the Mind, Cambridge University Press."},{"key":"ref_17","unstructured":"Newell, A. (1990). Unified Theories of Cognition, Harvard University Press."},{"key":"ref_18","unstructured":"Buckmann, M., Nguyen, Q.A., and Hill, E. (2025). Revealing economic facts: LLMs know more than they say. arXiv."},{"key":"ref_19","unstructured":"Zeng, W., Kurniawan, D., Mullins, R., Liu, Y., Saha, T., Ike-Njoku, D., Gu, J., Song, Y., Xu, C., and Zhou, J. (2025). ShieldGemma 2: Robust and tractable image content moderation. arXiv."},{"key":"ref_20","unstructured":"Superego GitHub (2025, July 16). Superego-Agent LGDemo (Branch: Fastapi_Mcp). GitHub. Available online: https:\/\/github.com\/Superego-Agent\/superego-lgdemo\/tree\/fastapi_mcp."},{"key":"ref_21","unstructured":"Mazeika, M., Phan, L., Yin, X., Zou, A., Wang, Z., Mu, N., Sakhaee, E., Li, N., Basart, S., and Li, B. (2024). HarmBench: A standardized evaluation framework for automated red teaming and robust refusal. arXiv."},{"key":"ref_22","unstructured":"Andriushchenko, M., Souly, A., Dziemian, M., Duenas, D., Lin, M., Wang, J., Hendrycks, D., Zou, A., Kolter, Z., and Fredrikson, M. (2025, January 5\u20139). AgentHarm: A benchmark for measuring harmfulness of LLM agents. Proceedings of the International Conference on Learning Representations (ICLR 2025), Vienna, Austria."},{"key":"ref_23","unstructured":"Pan, A., Chan, J.S., Zou, A., Li, N., Basart, S., Woodside, T., Ng, J., Zhang, H., Emmons, S., and Hendrycks, D. (2023, January 23\u201329). Do the rewards justify the means? Measuring trade-offs between rewards and ethical behavior in the Machiavelli benchmark. Proceedings of the International Conference on Machine Learning (ICML), Honolulu, HI, USA."},{"key":"ref_24","doi-asserted-by":"crossref","unstructured":"Shankar, S., Zamfirescu-Pereira, J.D., Hartmann, B., Parameswaran, A.G., and Arawjo, I. (2024). Who validates the validators? Aligning LLM-assisted evaluation of LLM outputs with human preferences. arXiv.","DOI":"10.1145\/3654777.3676450"},{"key":"ref_25","unstructured":"Vijil, Inc. (2025, May 20). Vijil Test Library: Evaluating LLM Trustworthiness Across Eight Dimensions. Available online: https:\/\/docs.vijil.ai\/tests-library\/index.html."},{"key":"ref_26","unstructured":"AI Safety Institute (2025, May 20). INSPECT: An Extensible Toolkit for AI Behavior Evaluation. Available online: https:\/\/inspect.aisi.org.uk."},{"key":"ref_27","unstructured":"Nasim, I. (2025, July 18). Governance in Agentic Workflows: Leveraging LLMs as Oversight Agents. OpenReview. Available online: https:\/\/openreview.net\/forum?id=fP02TFDJh8."},{"key":"ref_28","unstructured":"Zhang, J., Elgohary, A., Magooda, A., Khashabi, D., and Van Durme, B. (2024). Controllable Safety Alignment: Inference-Time Adaptation to Diverse Safety Requirements. arXiv."},{"key":"ref_29","unstructured":"Krishna, K., Cheng, J.Y., Maalouf, C., and Gatys, L.A. (2025). Disentangled Safety Adapters Enable Efficient Guardrails and Flexible Inference-Time Alignment. arXiv."},{"key":"ref_30","doi-asserted-by":"crossref","unstructured":"Wang, P., Zhang, D., Li, L., Tan, C., Wang, X., Ren, K., Jiang, B., and Qiu, X. (2024). InferAligner: Inference-Time Alignment for Harmlessness through Cross-Model Guidance. arXiv.","DOI":"10.18653\/v1\/2024.emnlp-main.585"},{"key":"ref_31","unstructured":"Ji, X., Ramesh, S.S., Zimmer, M., Bogunovic, I., Wang, J., and Bou Ammar, H. (2025). Almost Surely Safe Alignment of Large Language Models at Inference-Time. arXiv."},{"key":"ref_32","unstructured":"Li, X., Uehara, M., Su, X., Scalia, G., Biancalani, T., Regev, A., Levine, S., and Ji, S. (2025). Dynamic Search for Inference-Time Alignment in Diffusion Models (DSearch). arXiv."},{"key":"ref_33","unstructured":"Sharma, M., Tong, M., Mu, J., Wei, J., Kruthoff, J., Goodfriend, S., Ong, E., Peng, A., Agarwal, R., and Anil, C. (2025). Constitutional Classifiers: Defending against Universal Jailbreaks across Thousands of Hours of Red Teaming. arXiv."},{"key":"ref_34","unstructured":"Bai, Y., Kadavath, S., Kundu, S., Askell, A., Kernion, J., Jones, A., Chen, A., Goldie, A., Mirhoseini, A., and McKinnon, C. (2022). Constitutional AI: Harmlessness from AI Feedback. arXiv."},{"key":"ref_35","unstructured":"Bai, Y., Jones, A., Ndousse, K., Askell, A., Chen, A., DasSarma, N., Drain, D., Fort, S., Ganguli, D., and Henighan, T. (2022). Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback. arXiv."},{"key":"ref_36","unstructured":"Baker, B., Huizinga, J., Madry, A., Zaremba, W., Pachocki, J., and Farhi, D. (2025, March 12). Monitoring Reasoning Models for Misbehavior and the Risks of Promoting Obfuscation. OpenAI. Available online: https:\/\/openai.com\/index\/chain-of-thought-monitoring."},{"key":"ref_37","unstructured":"Greenblatt, R., Shlegeris, B., Sachan, K., and Roger, F. (2024). AI Control: Improving Safety Despite Intentional Subversion. arXiv."},{"key":"ref_38","unstructured":"Edelman, J., and Klingefjord, O. (2025, March 12). OpenAI x DFT: The First Moral Graph. Meaning Alignment Institute. Available online: https:\/\/meaningalignment.substack.com\/p\/the-first-moral-graph."},{"key":"ref_39","unstructured":"Edelman, J., and Klingefjord, O. (2025, July 18). Model Integrity. Meaning Alignment Institute. Available online: https:\/\/meaningalignment.substack.com\/p\/model-integrity."},{"key":"ref_40","doi-asserted-by":"crossref","first-page":"389","DOI":"10.1038\/s42256-019-0088-2","article-title":"The Global Landscape of AI Ethics Guidelines","volume":"1","author":"Jobin","year":"2019","journal-title":"Nat. Mach. Intell."},{"key":"ref_41","unstructured":"Beurer-Kellner, L., and Fischer, M. (2025, March 12). WhatsApp MCP Exploited: Exfiltrating Your Message History via MCP. Invariant Labs Blog. Available online: https:\/\/invariantlabs.ai\/blog\/whatsapp-mcp-exploited."},{"key":"ref_42","unstructured":"Beurer-Kellner, L., and Fischer, M. (2025, March 12). MCP Security Notification: Tool Poisoning Attacks. Invariant Labs Blog. Available online: https:\/\/invariantlabs.ai\/blog\/mcp-security-notification-tool-poisoning-attacks."},{"key":"ref_43","unstructured":"Betley, J., Tan, D., Warncke, N., Sztyber-Betley, A., Bao, X., Soto, M., Labenz, N., and Evans, O. (2025). Emergent misalignment: Narrow finetuning can produce broadly misaligned LLMs. arXiv."},{"key":"ref_44","unstructured":"Mowshowitz, Z. (2025, January 26). On Emergent Misalignment. Don\u2019t Worry About the Vase. Available online: https:\/\/thezvi.substack.com\/p\/on-emergent-misalignment."},{"key":"ref_45","unstructured":"Lu, C. (2025, January 26). Model Plurality. Combinations Magazine. Available online: https:\/\/www.combinationsmag.com\/model-plurality\/."},{"key":"ref_46","unstructured":"Lu, C., and Van Kleek, M. (2025, July 18). Model Plurality: A Taxonomy for Pluralistic AI. OpenReview. Available online: https:\/\/openreview.net\/forum?id=kil2mabTqx."},{"key":"ref_47","unstructured":"White, I., Nottingham, K., Maniar, A., Robinson, M., Lillemark, H., Maheshwari, M., Qin, L., and Ammanabrolu, P. (2025). Collaborating Action by Action: A Multi-agent LLM Framework for Embodied Reasoning. arXiv."},{"key":"ref_48","unstructured":"Yang, Z., Zhang, Z., Zheng, Z., Jiang, Y., Gan, Z., Wang, Z., Ling, Z., Chen, J., Ma, M., and Dong, B. (2024). OASIS: Open Agent Social Interaction Simulations with One Million Agents. arXiv."},{"key":"ref_49","unstructured":"Altera, A.L., Ahn, A., Becker, N., Carroll, S., Christie, N., Cortes, M., Demirci, A., Du, M., Li, F., and Luo, S. (2024). Project Sid: Many-agent simulations toward AI civilization. arXiv."}],"container-title":["Information"],"original-title":[],"language":"en","link":[{"URL":"https:\/\/www.mdpi.com\/2078-2489\/16\/8\/651\/pdf","content-type":"unspecified","content-version":"vor","intended-application":"similarity-checking"}],"deposited":{"date-parts":[[2025,10,9]],"date-time":"2025-10-09T18:19:13Z","timestamp":1760033953000},"score":1,"resource":{"primary":{"URL":"https:\/\/www.mdpi.com\/2078-2489\/16\/8\/651"}},"subtitle":[],"short-title":[],"issued":{"date-parts":[[2025,7,30]]},"references-count":49,"journal-issue":{"issue":"8","published-online":{"date-parts":[[2025,8]]}},"alternative-id":["info16080651"],"URL":"https:\/\/doi.org\/10.3390\/info16080651","relation":{},"ISSN":["2078-2489"],"issn-type":[{"type":"electronic","value":"2078-2489"}],"subject":[],"published":{"date-parts":[[2025,7,30]]}}}