Securing Autonomous Multi-Agent AI Systems: A Zero-Trust Framework for Threat Detection, Behavioral Monitoring, and Resilience in Enterprise Environments

This is an important issue in developing multi-agent systems in the context of distributed, autonomous, large-language models. Large-language model (LLM) agents delegate tasks, call tools and act on real-world agents. Generally, the most dangerous inputs arrived as natural-language messages from trusted peers, and one compromised agent could pass along its nefarious instructions to others. This paper proposes ZT-MAS, which brings together cryptographic agent identity and attestation, intent-bound capability tokens on every request, provenance tagging for taint propagation, a behavioural trust score that combines policy, sequence, provenance and drifting with taint propagation, a graduated response layer with step-up approval, restriction, quarantine and rollback, and a graduate response layer with step-down approval, restriction, quarantine and rollback. The framework is based on NIST SP 800-207, OWASP Top 10 for Agentic Applications, and MAESTRO threat model. We simulate an eight-agent enterprise workflow and sampled simulated a simulated a seeded simulation of a small company to determine the level of trust associated with goal hijacking, confused-deputy privilege escalation, prompt infection, memory poisoning, memory poisoning, and data exfiltration (500 episodes per cell). The perimeter zero-trust failed in all but behavioural tracking alone stopped every episode of a confused-deputy escalation, and behavioural monitoring stopped only volume-based exfiltration. While the perimeter and static zero-trust baselines succeeded in 99.6–100% of the episodes, ZT-MAS reduced this to 0–10.2% and blocked or held 2.1% of benign high-impact actions for approval. Many of the failures were injections manipulated to reflect the user’s real plan, and infection spread did not stop even when the user blocked the attack.

Authors

Institutions

Publication Details

Journal
International Journal of Scientific Research in Science and Technology
Published
2026-11-15
DOI
https://doi.org/10.32628/ijsrst2613620
Primary Topic
Access Control and Trust
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Securing Autonomous Multi-Agent AI Systems: A Zero-Trust Framework for Threat Detection, Behavioral Monitoring, and Resilience in Enterprise Environments

Sidhanta Panigrahy, Gokul Selvam Sadasivam
International Journal of Scientific Research in Science and Technology
Access Control and Trust
article

Securing Autonomous Multi-Agent AI Systems: A Zero-Trust Framework for Threat Detection, Behavioral Monitoring, and Resilience in Enterprise Environments

Sidhanta Panigrahy, Gokul Selvam Sadasivam
article en

Abstract

This is an important issue in developing multi-agent systems in the context of distributed, autonomous, large-language models. Large-language model (LLM) agents delegate tasks, call tools and act on real-world agents. Generally, the most dangerous inputs arrived as natural-language messages from trusted peers, and one compromised agent could pass along its nefarious instructions to others. This paper proposes ZT-MAS, which brings together cryptographic agent identity and attestation, intent-bound capability tokens on every request, provenance tagging for taint propagation, a behavioural trust score that combines policy, sequence, provenance and drifting with taint propagation, a graduated response layer with step-up approval, restriction, quarantine and rollback, and a graduate response layer with step-down approval, restriction, quarantine and rollback. The framework is based on NIST SP 800-207, OWASP Top 10 for Agentic Applications, and MAESTRO threat model. We simulate an eight-agent enterprise workflow and sampled simulated a simulated a seeded simulation of a small company to determine the level of trust associated with goal hijacking, confused-deputy privilege escalation, prompt infection, memory poisoning, memory poisoning, and data exfiltration (500 episodes per cell). The perimeter zero-trust failed in all but behavioural tracking alone stopped every episode of a confused-deputy escalation, and behavioural monitoring stopped only volume-based exfiltration. While the perimeter and static zero-trust baselines succeeded in 99.6–100% of the episodes, ZT-MAS reduced this to 0–10.2% and blocked or held 2.1% of benign high-impact actions for approval. Many of the failures were injections manipulated to reflect the user’s real plan, and infection spread did not stop even when the user blocked the attack.

International Journal of Scientific Research in Science and TechnologyVol. 13(6)
University of California, Berkeley (US)
Access Control and Trust
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon
Securing Autonomous Multi-Agent AI Systems: A Zero-Trust Framework for Threat Detection, Behavioral Monitoring, and Resilience in Enterprise Environments — Sidhanta Panigrahy, Gokul Selvam Sadasivam · International Journal of Scientific Research in Science and Technology (2026) | TGRS Research Map | TGRS