Mitigating Cascading Prompt Injections in Multi-Agent Systems via Semantic Boundary Mediation and Observability

This paper has presented a novel approach to securing multi agent AI swarms through the Semantic Boundary Frame work (SBF). By shifting the security focus from model layer alignment to system-layer Intent Verification, we have demonstrated that it is possible to mitigate Cascading Prompt Injections with high precision and minimal latency. Our GCP native implementation achieved a 74% increase in detection of unauthorized privilege escalations, proving that vector-based congruency checks are a viable defense against the “Confused Deputy” problem. While the SBF provides a robust defense for inter-agent communication, several avenues for future research remain:   Multi-Modal Intent Drift: Expanding the framework to analyze “Intent Drift” in multi-modal agents that process images and sensor data (critical for the automotive sector). ·        Decentralized Verification: Exploring the use of Blockchain-based notarization for agentic intent, creating a decentralized and immutable ledger of A2A instructions. ·        Adaptive    Thresholding:   Implementing Reinforcement Learning (RL) to automatically tune the drift threshold τ based on the evolving threat landscape of the specific organization.

1.       L. Lupinacci, J. Smith, and A. Gomez, “Prompt injection and inter-agent vulnerabilities in autonomous swarms,” IEEE Trans. Dependable Secure Comput., vol. 21, no. 2, pp. 112–125, 2025. 2.       M. Patel and K. Singh, “Model Context Protocol and semantic bound aries for AI orchestration,” in Proc. ACM CCS, 2025, pp. 210–219. 3.       [R. Wang and E. Clark, “Defense evolution for prompt injection in multi agent systems,” in Proc. IEEE S&P Workshop on AI Security, 2025, pp. 45–54. 4.       S. Chen and P. Zhao, “Vector-based intent drift detection for cloud-native agents,” IEEE Cloud Comput., vol. 9, no. 4, pp. 78–87, 2025. 5.       A. Miller, “Agent-to-agent protocols and trust models for autonomous AI,” AI Magazine, vol. 46, no. 1, pp. 34–42, 2025. 6.      N. Olzak, “Hidden instruction prompt injection techniques,” arXiv:2601.00001, 2026 7.      S. Gulyamov, “Ghost prompts and semantic camouflage in prompt injection,” in Proc. Euro S&P, 2026, pp. 99–108. 8.      Google Cloud, “Vertex AI Agent Builder security best practices,”            2025.            [Online].            Available: https://cloud.google.com/vertexai/docs/agent-builder/security