Search Results (968)

Prompt specifications for multi-agent large language model (LLM) systems carry data contracts and integration logic across interdependent files but are rarely subjected to structured-inspection rigor. We report a single-system case study of iterative, agent-driven auditing applied to AEGIS (Autonomous Engineering Governance and Intelligence System), a seven-lane production pipeline whose 7152-line specification surface was audited across nine rounds, surfacing 51 consistency defects (per-round counts of 15, 8, 12, 2, 8, 1, 4, 1, 0). We present a seven-category post hoc taxonomy with explicit coding rules, non-monotonic convergence consistent with cascading edits and audit-scope expansion, and a locked audit protocol. We further report two partial replications on a public synthetic mini-specification: a cross-LLM panel of four frontier vendors (OpenAI, Anthropic, Google, xAI; 12 traces; multi-vendor union detects all five seeded defects) and an inter-rater reliability check on a stratified subsample (Cohen’s $\kappa$ = 0.80 on category, 0.46 on severity). The full reproducibility bundle accompanies the submission. Full article

Post-Quantum Cryptography (PQC) migration to National Institute of Standards and Technology (NIST) Federal Information Processing Standards (FIPS) 203, 204, and 205 under the National Security Agency (NSA) Commercial National Security Algorithm Suite (CNSA) 2.0 is a multi-year, multi-domain transformation across cloud, enterprise, embedded, operational technology (OT), tactical, and national-security systems. Anthropic’s Claude Mythos Preview (April 2026) introduces artificial intelligence (AI)-accelerated cybersecurity capabilities that intersect this migration directly, performing autonomous reasoning against previously unknown vulnerabilities in production software—a qualitative departure from signature-based and static and dynamic application security testing (SAST/DAST) tooling. Drawing on federal guidance from NIST, NSA, the Office of Management and Budget (OMB), and the Cybersecurity and Infrastructure Security Agency (CISA), and on independent analyses from the Centre for Emerging Technology and Security (CETaS) and the UK AI Security Institute, we present a lifecycle and architecture analysis of how Mythos-class models alter PQC migration timelines, risk surfaces, lifecycle dependencies, and architectural constraints. Modeling Mythos as both accelerator and destabilizer, we derive an analytic projection of a compressed two-to-four-year migration window for highest-exposure systems, against traditional baselines of five-to-ten years for small organizations and twelve-to-fifteen-plus years for large enterprises. The compression collapses human-labor bottlenecks in discovery, planning, and code modification, not cryptography itself. We propose a lifecycle-aligned migration model, an updated cost model, and governance requirements for frontier-model access. The binding constraint shifts domain-conditionally: defender capacity at adversary tempo governs software-analytical phases, while non-compressible external cadence governs embedded and regulated domains. Full article

Mechanistic interpretability seeks to reverse-engineer the computational circuits within large language models, but current methods rely on exhaustive or heuristic search over exponentially many feature interactions. This paper introduces Active Circuit Discovery (ACD), a framework that combines attribution-graph analysis with active inference to select interventions efficiently. ACD uses Anthropic’s circuit-tracer library as its attributiongraph backend, applying Edge Attribution Patching with transcoders to identify the active transcoder features for each prompt. A partially observable Markov decision process (POMDP) agent, implemented with pymdp, maintains a multi-factor generative model of feature importance, layer role, and causal influence. At each step, the agent selects both a target feature and an intervention type (ablation, activation patching, or feature steering) by minimising Expected Free Energy over the joint feature–action space, and it learns its observation model online through Dirichlet parameter updates. ACD is an interventionselection layer over existing attribution-graph tools; it is not a whole-circuit discovery method, and no claim of state-of-the-art circuit discovery is made. The framework is evaluated on Gemma-2-2B (26 layers) and Llama-3.2-1B (16 layers) across four settings: Indirect Object Identification (IOI), multi-step reasoning, feature steering, and a multidomain benchmark spanning geography, mathematics, science, logic, and history. With a budget of 20 interventions per prompt, an ablation-only agent scored by bounded oracle efficiency against the ablation oracle reaches 82.0% efficiency on Gemma IOI and 73.0% on Gemma multi-step. It exceeds random selection by 43.5% (relative) on Gemma IOI (paired permutation p = 0.031) and is competitive with greedy ranking, a heuristic UCB bandit, and a plain UCB baseline. A direct Edge-Attribution-Patching ranking is itself a strong baseline that the agent does not consistently surpass, and on Llama multi-step the agent reaches 9.3% efficiency (37.8% with finer layer-role bins). All comparisons report bootstrap 95% confidence intervals. The full multi-action agent is characterised separately by a Relative Cumulative KL, a steering-driven amplification factor reported apart from the bounded efficiency. Feature steering changes the top-1 prediction in a dose-dependent manner, but a matched random-feature control shows that circuit-selected features are only marginally, and not significantly, more steerable than random active features at large multipliers, indicating that part of the effect is generic activation scaling. Multi-domain analysis shows task-dependent circuit structure, with IOI circuits concentrated in late layers and reasoning and scientific knowledge recruiting early and middle layers. Code, notebooks (free T4), AMD64/aarch64 Docker images, and raw results are publicly available. Full article

The Maule Region in Chile possesses a rich cultural heritage associated with petroglyphs created by ancient hunter-gatherer inhabitants. This rock art has suffered damage over time due to natural and anthropic causes. Fostering positive attitudes toward petroglyphs may influence behavioral intentions related to their preservation. This study evaluates an augmented reality card game developed to promote positive attitudes toward the rock art heritage of the Maule Region, examining its usability and the effects of incorporating augmented reality elements. The game achieved a System Usability Scale (SUS) score of 79.7 (SD = 14.2), corresponding to an A-grade on the Sauro-Lewis curved grading scale, indicating good usability.Participants in the game condition showed higher heritage attitudes than controls (M = 6.13, SD = 0.80, $t_{\left(24\right)}$ = −2.33, p = 0.028). Augmented reality enhanced attitudes at moderate levels of usability (B = −1.02, p = 0.043), but produced no detectable main effect in mean comparisons alone. The results indicate that the game constitutes a system with adequate usability, effective in fostering positive attitudes toward cultural heritage, and that augmented reality enhances attitudinal outcomes under conditions of moderate perceived usability. Full article

Although prior research indicates that expert reviewers identify AI-generated academic texts with low accuracy, the quantitative analysis presented in this paper has revealed marked, measurable differences between human-authored and AI-generated works. We investigate this duality in the context of Hungarian as an under-represented training language: on one hand, we perform a quantitative text analysis of the lexical, syntactic, and stylistic features of Hungarian-language academic essays by human authors (doctoral candidates) and those generated by Google Gemini, OpenAI GPT, and Anthropic Claude models. On the other hand, using a blind experimental design, we analyze how human reviewers (N = 391) with varying levels of expertise perceive and assess the quality of the texts. The quantitative analysis showed that AI-generated essays are characterized by lower lexical diversity and an absence of epistemic markers. The human evaluation yielded complex results: reviewers active in academic practice (members of the academically active and academically passive clusters) acknowledged the formal and logical precision of the AI-generated texts, yet they noted a lack of originality and critical depth. Reviewers less engaged with academic practice (members of the non-academic and inactive clusters), in contrast, were primarily persuaded by the more natural style and originality of the human-authored texts. The findings suggest that with moderate-level prompting and the provision of source literature, an AI-generated essay can be created in a few hours that reviewers deem superior to human work in certain aspects, such as formal and logical precision. Furthermore, our findings suggest that with targeted, more sophisticated prompt engineering, the quality gap between AI-generated and human-authored texts could narrow further. These findings have significant implications for assessment methods in higher education and for the regulation of academic publishing. Full article

The Amazon Basin harbors a high diversity of phlebotomine sand flies, including several species that act as vectors of zoonotic pathogens such as Leishmania. Land-use changes, particularly forest conversion to agriculture, alter the sand fly diversity and community structure, with implications for the transmission of American Tegumentary Leishmaniasis (ATL). We evaluated the effects of forest-to-agriculture conversion on sand fly diversity and species composition in two rural areas, on opposite sides of the Amazonas River, in the Brazilian Amazon Region. Sand flies were collected using Center of Disease Control (CDC) light traps, installed in the forest and cropland environments, at the Presidente Figueiredo (North) and Urucurituba (South) municipalities, in Amazonas, Brazil. We collected a total of 1778 phlebotomine sand flies from 15 genera and 69 species. The most abundant species were Micropygomya rorotaensis (n = 436; 24.52%), Nyssomyia antunesi (n = 297; 16.70%), Sciopemyia sordellii (n = 101; 5.68%), Bichromomyia flaviscutellata (n = 84; 4.72%) and Evandromyia monstruosa (n = 72; 4.04%). In addition, four sand fly species were recorded for the first time in the Amazonas state: Brumptomyia mesai, Pressatia calcarata, Evandromyia aldafalcaoae and Lutzomyia carvalhoi. Sand fly richness, diversity, and community composition varied between riversides and environments, reflecting strong effects of anthropogenic disturbance. Although croplands supported reduced and more heterogeneous assemblages, several medically important vector species persisted across both environments. Species turnover was high, but patterns of species loss were weak, suggesting that community reorganization was driven by non-directional compositional change process. Our results indicate that land-use change reshapes sand fly communities without eliminating disease vectors, potentially increasing ATL transmission risk at the forest–anthropic interface. Full article

Background: Data extraction for systematic reviews is highly resource-intensive. This study evaluated four frontier large language models (LLMs) on complex structured metadata extraction from specialized neuroimaging artificial intelligence (AI) literature to determine their performance in automated evidence synthesis. Methods: We compared Google Gemini 3 Pro Preview, Anthropic Claude Opus 4.5, Perplexity Sonar Pro, and OpenAI GPT 5.2. Using a standardized prompt, each model extracted 22 variables from 91 peer-reviewed neuroimaging AI articles. The variables were stratified into low-, medium-, and high-complexity tiers. The performance was measured via the exact-match accuracy against a consensus-based expert ground truth. Results: The overall exact-match accuracy was moderate. Gemini 3 Pro Preview achieved the highest overall rate (56.4%), followed by Sonar Pro (52.1%), Claude Opus 4.5 (51.3%), and GPT 5.2 (46.5%). Gemini significantly outperformed all other models (p < 0.001). The performance declined dramatically as the variable complexity increased. Across models, the accuracy was 88.9–92.9% for low-complexity categorical fields, 47.0–63.3% for medium-complexity text extraction, and 2.7–15.5% for high-complexity variables requiring clinical judgment or multi-section synthesis. The most common type of error was misclassification. All four models scored 0% on the main performance metric, but this reflected a representational mismatch with the ground truth rather than extraction failure, indicating that the exact-match accuracy underestimates the true semantic performance. Conclusions: Frontier LLMs can effectively automate the retrieval of simple categorical data, but have serious difficulties with methodological variables that are complex. Although extraction can be fully automated for low-complexity fields, human review remains essential for context-dependent variables that require clinical judgment. Full article

Vegetation spatial heterogeneity is fundamental to biodiversity management and ecosystem service provision, yet detailed phytosociological mapping of forest vegetation remains largely unresolved in the Colombian Orinoquia. This study characterized the geographic distribution of forest vegetation through the integration of 178 field surveys, environmental complex variables defined by geomorphological and bioclimatic gradients, and multi-sensor satellite imagery combining Landsat-8 optical bands and Sentinel-1 dual-polarization data, processed within a Random Forest classification framework in Google Earth Engine. Classifications achieved overall accuracies between 0.910 and 0.975 and Kappa coefficients above 0.93, identifying 24 phytosociological alliances or geobotanical formations distributed across approximately 7,565,696 ha, representing 34.63% of the region. Forest cover ranges from 10.95% in the Floodplain to 55.22% in La Macarena, with the High Plain concentrating the greatest formation diversity. The spatial organization of forest vegetation is primarily governed by the geomorphological gradient—fluvial, denudational, and structural—and limiting bioclimatic factors, together with their associated edaphic−hydrological regimes, with anthropic transformation driven by cattle ranching and agricultural expansion constituting the principal threat to forest cover. These results advance beyond existing land cover surrogates, providing an empirically validated cartographic framework for biodiversity assessment, habitat modeling, and natural capital management in the Colombian Orinoquia. Full article

Anthropic’s April 2026 Claude Mythos Preview release established a new operational threat category: frontier AI systems whose extended-context reasoning, recursive self-correction, native system-tool integration, and agentic scaffolding render dominant AI safety paradigms—RLHF, output filtering, contractual access vetting, human-in-the-loop supervision—insufficient as sole controls. This paper develops a defense-in-depth reference architecture against that category, structured around four named contributions: a five-indicator operational definition of the Mythos-class (capability conjoined with scaffold, access pattern, autonomy depth, and persistence); the Mythos-Class Posture Rubric (MCPR), a three-tier detection framework spanning evaluation, deployment, and runtime with explicit routing to mitigation layers; a four-layer mitigation stack comprising the Vetted-Access Operational Pattern (VAOP), Authority-Bound Output Release (ABOR) cryptographically grounded in FIPS 203/204/205 post-quantum primitives, and the Compute-Plane Isolation Profile (CPIP); and an integrated architecture that crosswalks to the NIST AI Risk Management Framework, NIST Cybersecurity Framework 2.0, and CISA Zero Trust Maturity Model 2.0. The architecture is applied to three deployment surfaces—post-quantum cryptography migration, federal AI supply-chain assurance, and critical-infrastructure operational technology defense—demonstrating that the four contributions generalize across heterogeneous operational contexts. The contribution is a reference design rather than a deployed system; limitations, falsifiability criteria, and a research agenda for empirical refinement are developed. Full article

The activity of an Apennine wolf (Canis lupus italicus) family inhabiting the natural reserve Calanchi di Atri in central Italy was monitored during the post-reproductive period (May–October) of two consecutive years (2023–2024), using ten camera trap sites. Detections were classified into adults and pups. Although records cover a limited period and focus on a single pack, they allowed the detection of variations in the spatiotemporal activity of the wolf family. In the first year, wolf activity peaked in summer, with adults frequently supervising pups at rendezvous sites. In the second year, activity by both adults and pups declined significantly and was accompanied by an evident shift in territory use. In addition to potential intrinsic factors, such as individual variability and litter dynamics, these variations may also reflect increased environmental stressors and anthropogenic disturbance. These findings provide insights into how wolves adapt their behavior in human-modified landscapes and highlight the importance of integrating human–wildlife dynamics into conservation and management strategies. Full article

The degradation of forest landscapes in West Africa, particularly in Togo, threatens ecological and socio-economic sustainability. This study analyzes the spatio-temporal dynamics of land use in the central plains of Togo between 1991 and 2022, and projects its evolution for 2030 and 2050 to guide restoration strategies. The methodology integrates the interpretation of Landsat images (1991, 2005, 2022) and the analysis of indicators, including conversion rates and the anthropization index. Prospective modeling (Markov chains and neural networks) follows a trend scenario. The results reveal a sharp decline in natural forest formations: dense semi-deciduous and dense dry forests (−50.55%) and woodlands (−62.06%), converted mainly to cropland, plantations, and built-up areas. Shrub/tree savannas, the dominant class, represent a transitional stage resulting from forest degradation. The average annual deforestation rate is 0.75%. The ecological disturbance index increased from 0.24 (1991) to 0.45 (2005), and then to 0.56 (2022), reflecting increased human impact and fragmentation. Projections indicate that these trends will continue, highlighting the growing vulnerability of ecosystems and the need to integrate this dynamic into sustainable management and restoration policies. Full article

Background/Objectives: Antimicrobial resistance (AMR) is a global threat to human, animal, and environmental health. Among bacteria, E. coli is frequently used as a key indicator of AMR. Despite their genetic proximity to humans, studies on AMR in Non-Human Primates (NHPs) remain limited, particularly in semi-anthropized environments. This study aims to characterize the antibiotic resistance profiles and phylo-groups of E. coli isolated from NHPs and humans at a primatology center. Methods: A total of 143 stool samples were collected, including 125 from NHPs and 18 from humans. Isolates were cultured on Eosin Methylene Blue agar and then identified by MALDI-TOF mass spectrometry. Antibiotic susceptibility was assessed using the Kirby–Bauer disk diffusion method, with 30 antibiotics following CASFM-EUCAST recommendations. E. coli phylo-groups were characterized by quadruplex PCR according to the Clermont method, targeting the genes. Results: A total of 122 E. coli isolates (85.31%) were recovered, with comparable prevalence observed across NHPs and human staff. More than half of the isolates (55.74%) were resistant to at least one antibiotic tested, and 12.3% were classified as multi-drug resistant (MDR). Resistance rates of isolates in Mandrillus sphinx, Pan troglodytes, and humans were 50.6%, 57.7%, and 80.0%, respectively, with no significant statistical difference (p = 0.11). A single Extended-Spectrum Beta-Lactamase (ESBL) producing isolate was identified in the mandrill. Phylo-group analysis revealed the dominance of group A (50%), followed by groups B1, D, and C. Conclusions: Resistance profiles and phylo-group distribution among NHPs could suggest bacterial exchange and potential for cross-transmission of AMR within the shared environment. Full article

The increasing consumption of fish has raised public health concerns regarding fish-borne zoonotic helminths (FBZHs), which are recognized as significant food-borne parasites worldwide. In freshwater environments, Clinostomum complanatum, Opisthorchis felineus, Pseudamphistomum truncatum, Dibothriocephalus latus and Eustrongylides excisus are of particular concern in Italy and neighbouring countries. This study aimed to assess the prevalence of these FBZHs in five commercially and ecologically relevant freshwater fish species from Lake Varese, a heavily anthropized and understudied basin in northern Italy. A total of 59 fish were examined via necropsy and stereomicroscopic inspection of skeletal muscles. Only Eustrongylides spp. larvae were detected, with a prevalence of 16.9%. Molecular analysis (ITS region) identified them as E. excisus. This study reports, for the first time in Western Europe, E. excisus in Sander lucioperca (p = 12.5%) and Esox lucius (p = 8.3%). The highest prevalence occurred in Silurus glanis (p = 37.5%), followed by Perca fluviatilis (p = 25.0%), while Tinca tinca showed no infection. These findings confirm that among the FBZHs considered, E. excisus is currently present and expanding both in host range and geography in Italian lakes, underscoring the need for updated epidemiological data to support risk assessment, food safety and zoonotic parasite control in freshwater fisheries. Full article

Despite the widespread perception of Chilean Patagonia as water-abundant, the region exhibits marked climatic and landscape heterogeneity. This study evaluates relative water availability across Coyhaique Province (12,712 km²), where projections indicate a trend toward warmer and drier conditions. The province has a marked west–east gradient: humid valleys in the west contrast with much drier areas to the east, where most of the population and development are concentrated. To identify water stress hotspots, we combine spatially explicit water yield estimates derived from the InVEST Seasonal Water Yield model with an anthropization index used as a proxy for water demand, constructing a relative Water Stress Index. The results indicate that water stress increases toward the east, driven by the combined influence of climate variables and anthropogenic pressure. These results indicate that the characterization of Patagonia as uniformly water-rich does not hold at the provincial scale, and highlight the limitations of coarse regional assessments in capturing intra-regional hydrological heterogeneity. The spatial pattern of water stress revealed here exposes a mismatch between the resolution at which hydrological heterogeneity operates and the scale at which prevailing water governance frameworks are formulated, underscoring the need for bottom-up, fine-resolution diagnostics that incorporate local hydrological variability into water planning and governance. The province-scale analysis presented here provides a representative case study for Aysén and illustrates the broader relevance of spatially explicit diagnostics in contexts where regional indicators mask local water stress. Strengthening monitoring networks, protecting headwater catchments, and promoting a decentralized approach to water management remain essential to reduce the risk of human-driven water scarcity. Full article

Tailings are unconventional geomaterials that require dynamic characterisation due to seismic hazards at several storage facilities. Due to the anthropic origin of these materials, their dynamic properties differ from those reported for natural soils. In particular, the damping ratio is a relevant parameter that controls the dynamic response of tailings storage facilities. It can be estimated using different experimental methods. The objective of this research is to disclose the results obtained through laboratory tests in which the damping ratio was evaluated independently by Half-Power Bandwidth or the free-vibration decay methods. A comprehensive testing plan comprising resonant column tests and free-vibration decay tests was carried out on three types of tailings from the Riotinto mines (Huelva, Spain): Cerro Salomón Sand (CSS), High-Density Sludge (HDS), and Copper Lamas (CL). These tests were carried out under different effective consolidation pressures and torsional excitations. The results allowed the establishment of a series of relationships between the testing conditions and the identification of differences between the methods for tailings. Full article