
Large Language Model based support systems are increasingly used in many real world applications. These systems remain vulnerable to attacks such as prompt injection, tool hijacking, memory poisoning, RAG poisoning and multi agent infection. Also communication between agents may be exposed to harvest-now-decrypt-later (HNDL) attacks from future quantum adversaries. Existing solutions generally address only individual threats and do not provide a unified framework for securing both the LLM workflow and inter agent communication. Our paper presents SentinelX AI, a post quantum secured multi agent defense framework that examines each request before it reaches the LLM. Our developed framework consists of six processing stages including a Trust Score (TS) based risk assessment module, a blacklist based Memory Sanitizer, a controlled Attack Simulator and an Execution Gate for decision making. To protect communication between agents where SentinelX AI employs ML-KEM-512 (NIST FIPS 203) for key establishment and AES-256-GCM for authenticated encryption at every agent boundary.SentinelX AI reduces the Attack Success Rate from 73.0% to 4.0%, achieves an Attack Detection Rate of 96.0% with zero false positives, maintains a Task Success Rate of 91.3% and yields a Defense Effectiveness Score (DES) of 0.876, while introducing an average post quantum communication overhead of only 3.71 ms per agent hop (0.40% of the LLM inference time). SentinelX AI enhances the security of multi agent LLM systems with minimal additional computational cost.
Distributed and targeted attacks demand Network Intrusion Detection Systems (NIDS) capable of cooperating without exposing sensitive data. Although Federated Learning (FL) enables model training without centralizing raw traffic, sample-proportional aggregation may dilute a minority attack signal observed by only one organization. This work proposes Asymmetric Anomaly-aware Immune Aggregation (A2IA), a divergence-sensitive strategy for federated heterogeneous graph neural network (HGNN)-based intrusion detection, oriented towards amplifying updates associated with rare attacks. A two-layer Heterogeneous Graph Attention Network (HAN) is adopted as the local backbone for the proof of concept, although the aggregation mechanism operates on model updates and is not intrinsically tied to HAN. The mechanism automatically estimates a consensus reference from the closest pair of projected client updates, without prior knowledge of which client has observed attack traffic. A controlled single-round proof of concept on CICIDS2017 was evaluated with three honest clients and five random seeds. FedAvg obtained a mean attack recall of 0.0000, whereas A2IA preserved a mean recall of 0.9500 ± 0.0944, with a mean precision of 1.0000. These results support the preservation of a client-localized minority signal in the evaluated setting.
Digital forensic investigations are becoming increasingly challenging as cybercriminals exploit virtual environments and privacy-enhanced browser features to conceal malicious activities. This study investigates Mozilla Firefox’s private browsing mode inside virtual environments and the recoverability of digital evidence. Four controlled scenarios were examined: browsing-only activity, file download and deletion, post-reboot conditions, and network-level analysis. Live memory, disk, and network analysis were conducted using FTK Imager, Autopsy, Volatility 3, and Wireshark. The findings demonstrate that URLs, search queries, and traces of downloaded files remain recoverable despite the use of private browsing. The study shows that private browsing provides limited local privacy and highlights the importance of combining RAM, disk, and network analysis when investigating private browsing activity in virtualized environments.
Federated learning (FL) enables collaborative training without centralizing sensitive data, yet real-world deployment remains hard due to scalability bottlenecks, infrastructure complexity, communication overhead, and limited robust aggregation. We present a lightweight, modular, production-ready FL framework, and outline its design and vision: (1) a hierarchical three-tier architecture (Orchestrator–Coordinator–Client) with push-based task streaming for horizontal scalability; (2) an extensible library of 10+ aggregation algorithms across six families, including Byzantine-robust and adaptive methods; and (3) a real-time interactive dashboard for configuration, monitoring, and comparison. We report preliminary results showing effective horizontal scaling, and describe ongoing work toward a full system evaluation.
The rapid advancement of Generative Artificial Intelligence (GenAI) is transforming how organizations process information, generate insights, and support enterprise decision-making. While AI systems can analyze large volumes of structured and unstructured data with unprecedented speed, effective decision-making continues to depend on human judgment, domain expertise, ethical reasoning, and contextual understanding. This paper proposes a Human-Centered Collaborative Intelligence Framework that integrates the analytical capabilities of Generative AI with human expertise to support trustworthy enterprise decision-making. The framework positions humans and AI as complementary collaborators and emphasizes human oversight, transparency, accountability, and governance as fundamental principles for responsible AI adoption. The study adopts a design science approach, drawing on established literature in enterprise systems, trustworthy AI, and human-AI collaboration to develop the proposed framework. Its practical applicability is illustrated through representative enterprise decision scenarios spanning supply chain management, procurement, finance, customer operations, and business planning. The paper demonstrates how collaborative intelligence can improve decision quality, organizational agility, and governance while preserving human accountability. The proposed framework provides a structured foundation for integrating Generative AI into enterprise environments and offers practical guidance for organizations seeking to implement trustworthy human-AI collaboration. Future research directions are identified to support empirical validation across diverse enterprise settings.