Cloud Sentinel-AI: An Explainable Behavioral-Provenance and Self-Healing Security Framework for Autonomous AI Agents in Multi-Cloud Environments
Abstract
The rapid adoption of autonomous artificial intelligence (AI) agents in cloud environments has introduced new security challenges due to their ability to dynamically select tools, access heterogeneous resources, interpret external information, and modify execution plans at runtime. Conventional authentication and static authorization mechanisms are often insufficient to capture the evolving security context of such agent-driven activities. This study proposes CloudSentinel-AI, an explainable, behavior-aware, and self-healing security framework for protecting autonomous AI agents operating across multi-cloud environments. The proposed framework continuously constructs a behavioral provenance model that links agent identity, task context, data sources, tool invocations, API operations, cloud resources, and resulting actions. A hybrid detection mechanism combines behavioral analysis, contextual risk assessment, and provenance-deviation analysis to identify potentially malicious or unauthorized agent behavior. An explainable decision layer associates detected risks with supporting evidence and determines proportional security responses, including additional verification, capability restriction, credential revocation, and agent isolation. The proposed prototype architecture integrates Python, FastAPI, PostgreSQL-compatible storage, graph-oriented provenance representation, Docker-based deployment, a web-based monitoring dashboard, and a Flutter mobile client. To support reproducible evaluation, the CloudSentinel-AgentSec Dataset (CASD) is proposed as a synthetic benchmark containing normal and adversarial autonomous-agent sessions. The evaluation methodology considers precision, recall, F1-score, false-positive rate, detection latency, containment latency, and system resource overhead. No performance claims are made until the prototype is evaluated using the defined benchmark. The proposed work therefore provides a reproducible technical architecture and systematic evaluation methodology for improving the security, explainability, and resilience of autonomous AI agents in multi-cloud environments.