Dynamic multi-objective optimization problems (DMOPs) have attracted significant attention in recent years. However, some existing approaches do not fully exploit the geometric structure of the Pareto optimal front, or rely on relatively coarse distance-based matching, which may lead to less accurate solution alignment and degraded prediction performance under dynamic changes. To address these issues, this paper proposes a Topological Prior and Vector Migration-based Multi-Objective Evolutionary Algorithm (TPV-MOEA) to enhance sub-region partitioning and improve prediction accuracy. Specifically, TPV-MOEA incorporates a two-layer graph-based topological aggregation module to aggregate neighborhood information of shared points, enabling topology-aware ordering along the Pareto front manifold for more reliable sub-region partitioning. Within each sub-region, a vector migration strategy is employed to adaptively transfer global MOEA/D reference vectors into local regions, enabling better alignment with the underlying Pareto front structure and achieving more accurate association between non-dominated solutions and their historical counterparts. New populations are generated through direction-guided alignment and position prediction, thereby enabling effective adaptation to dynamic environmental changes. Comparisons with five state-of-the-art algorithms show that TPV-MOEA achieves superior convergence and diversity in tracking the Pareto front under dynamic environments.
Tool-using large language model (LLM) agents turn inputs into proposed external actions, making query screening and call authorisation distinct security decisions. We evaluate this distinction in a seven-tool mock sandbox with TinyLlama-1.1B, Phi-3.5-mini, and Mistral-7B across 3,735 generations covering direct attacks, legitimate tool tasks, and indirect injection under a policy-by-tool-exposure factorial. Routing-corrected rule-positive calls occurred in 63/315, 147/315, and 109/315 direct-attack generations, whereas legitimate-task completion was 96/210, 188/210, and 204/210, respectively. In a four-arm call-level comparison, a deterministic label-blind gate blocked 363/365 rule-positive calls and retained 398/630 legitimate task completions, against 488/630 with no defence and 20/630 with query screening alone. Blinded annotation of a targeted 157-row sample showed strong inter-rater agreement on tool-call presence (κ=0.8637) and unsafe-call judgement (κ=0.9215). The unified parser achieved recall 0.9259 against human judgements, versus 0.5556 for the legacy parser. The indirect factorial yielded 5 parsed calls in 2,160 generations, 4 of them from benign documents; it therefore supports exact event-rate bounds rather than a policy-effect estimate. Trained on source-confounded data, the screen flagged 67/70 harmless tool-requesting tasks but 0/1,207 benign general prompts, behaviour consistent with sensitivity to tool vocabulary. Within the evaluated setting, these results support a layered design in which learned screening provides optional triage and deterministic call-level mediation governs execution.
Metamaterials, engineered at the unit-cell level to manipulate wave propagation beyond the limits of conventional materials, rarely combine broadband airborne-sound absorption with structure-borne vibration attenuation in a single lightweight panel. This study proposes a two-dimensional, starfish-inspired multifunctional metamaterial addressing both transmission paths within a compact, mass-efficient unit cell. Its vibro-acoustic performance is characterised through thermoviscous finite element simulation, Bloch–Floquet dispersion analysis, and impedance-tube and shaker-based measurements on additively manufactured specimens. The measured panels achieve an absorption coefficient above 0.5 over a bandwidth of approximately 600 Hz within 400–1600 Hz, with near-unity absorption (α ≈ 0.9–1.0) at the resonant peaks, while dispersion analysis over 0–1200 Hz reveals multiple elastic bandgaps between approximately 610 and 1200 Hz, with measured accelerance reductions of up to 40 dB relative to the below-gap response. The starfish cell operates as a Helmholtz-type resonator, with the base radius R and arm amplitude A conventionally tuning the resonance through the cavity volume across approximately 440–1200 Hz. The distinguishing feature is the pore offset W, which shifts the peak by up to 200 Hz ( ∼ 20%) at constant cavity volume, depth, and neck geometry, by adding inertance without compliance through the arm channels a route unavailable in conventional Helmholtz or lattice-based absorbers; the arm number S affects the resonance only once W is non-zero. The interlocked arrangement yields a core mass per unit area approximately 48.3% lower than an equivalent honeycomb. These results position the design as a lightweight candidate for integrated noise and vibration control in aircraft, railway, and automotive structures.
This paper questions the form of power attributed to migrant women entrepreneurs often emanating from their pre-established social positions, which limits their engagements to mere reaction and adaptation to dominant power holders (i.e. specific groups, contexts, structures). Drawing on the life stories of four migrant women entrepreneurs of Turkish origin in Sweden and the Netherlands, we examine how they experience power as an embodied and affective phenomenon through a multiple case study approach. Going beyond the positional power attributed to women, we adopt a critical feminist perspective on power and demonstrate how migrant women entrepreneurs actively exercise different 'modalities of power' (power-over, power-to and power-with) and build 'power agility'. The study contributes to migrant women's entrepreneurship by uncovering their power as an emergent attribute rather than as a fixed parameter and by articulating how this 'agility' is at the core of migrant women entrepreneurs' capacity to generate transformative effects crossing over individual, relational, contextual, and systemic aspects and levels.
Autonomous AI agents powered by large language models (LLMs) with structured function-calling interfaces have greatly expanded capabilities for real-time data retrieval, computation, and multi-step orchestration. However, the rapid growth of plugins, connectors, and inter-agent protocols has outpaced security practices, leading to brittle integrations — plugin APIs and protocol adapters that rely on ad-hoc authentication, inconsistent schemas, and weak validation — making them vulnerable to failures and exploitation. This survey introduces a unified end-to-end threat model for LLM-agent ecosystems, spanning host-to-tool and agent-to-agent communications, and catalogs over thirty attack techniques across Input Manipulation, Model Compromise, System and Privacy Attacks, and Protocol Vulnerabilities. For each category, we provide a formal mathematical formulation of the underlying threat model, defining attacker capabilities, objectives, and affected layers to enable systematic analysis. Representative examples include Prompt-to-SQL (P2SQL) injections and the Toxic Agent Flow exploit in GitHub’s MCP server. For each category, we assess feasibility, review defenses, and outline mitigation strategies such as dynamic trust management, cryptographic provenance tracking, and sandboxed agentic interfaces. The framework was validated through expert review and cross-mapping with real-world incidents and public vulnerability repositories (e.g., CVE, NIST NVD) to ensure practical relevance. Compared to prior surveys, this work provides the first integrated taxonomy bridging input-level exploits and protocol-layer vulnerabilities in LLM-agent ecosystems while introducing formal system definitions for each threat class. Ultimately, it offers actionable insights for securing next-generation AI agents through layered defense and continuous verification. Our work provides a comprehensive reference to guide the design of secure and resilient LLM-agent workflows.
