Detailed Program

Local-inference language models structurally reduce and displace the human-supervision bottleneck that has historically constrained influence operations. We formalize SAINTS (Synthetic Autonomous Identity Network for Targeting with Saturation), a class of infrastructure in which each persona runs on a physically isolated node executing a model fine-tuned, at the weight level, on the persona’s behavioral profile rather than on a mission. Autonomy here denotes passive identity maintenance, not unsupervised active campaigning. SAINTS is presented as a formal threat model rather than an implemented or validated system : the contribution is a model precise enough that its breaking surfaces become analyzable. We characterize four architectural properties (non-correlation, embodied behavioral coherence, opaque action chains, saturation), formalize the cost asymmetry between deployment and investigation, and identify five detection surfaces the architecture cannot absorb. The central claim is a structural inequality : under the stated assumptions, investigation cost grows super-linearly in network size while deployment cost grows linearly, suggesting the possible existence of a threshold 𝑘* (conditional on the defender’s success function and budget) beyond which investigation becomes, in this model, economically irrational. The numerical estimates given here are illustrative, not normative, and require empirical calibration. Patience (waiting for human-operator failure, likely over long horizons) emerges as a principal defensive vector whose cost scales symmetrically with attacker deployment.

Post-quantum cryptography (PQC) migration starts with discovery, yet a flat inventory cannot connect observed certificates, libraries, pipelines, or devices to dependent services. CIGMA deterministically resolves heterogeneous observations into a typed dependency graph with locator-backed provenance for every accepted assertion. Its closed ontology, quality gate, and allow-listed paths form an executable structural contract. Impact, migration feasibility, and evidence confidence remain separate; policy tiers, Pareto fronts, and stable tie-breaks define the primary order, with a weighted comparator retained as a baseline. CIGMA-Lab supplies 46 entity and 34 relation observations from six source families; an oracle of 23 entities and 34 relations is held out in a separate evaluator-only tree. CIGMA attains 1.0 on all four reconstruction F1 metrics and NDCG@10 ≈ 0.90 against author-defined priorities. We also run 384 paired method runs over 128 deletion scenarios, a TFIDF-Cosine external-library comparator, and a synthetic scaling probe. In a separately sealed, fixed-input cohort of 30 gpt-5.6-luna calls, deterministic validation attains mean precision 1.0000, accepts no oracle-false triple in those calls, and raises mean candidate F1 from 0.9361 to 0.9554. These results establish reproducible behavior on one controlled fixture, not organizational accuracy or expert agreement. The submission artifact includes the separated oracle, executable baselines and ablations, offline model replay, and a two-phase protocol for sealing predictions for future independent replications before oracle access.

The RéSISTeCC exercise protocol is the result of three years of interdisciplinary work spanning incident response, crisis management, and cognitive ergonomics. The RéSISTeCC project stands for resilience through strategic and technical immersive simulation of cyber crises and was funded by NextGenerationEU and France 2030 through the DEFFINUM call for proposals. The latest version of this protocol offers a seven-hour cyber crisis training session designed for organizations with limited cybersecurity experience. Its purpose is to assess and test crisis management processes from both technical and strategic perspectives. The exercise protocol involves a strategic unit and a technical unit responsible for cyber crisis management, which are required to collaborate and coordinate their actions simultaneously. Only the technical unit is immersed in a cyber range simulating a generic "LOREM" information system. Both units receive phone calls from the facilitators and can call or visit one another. Traditionally, cyber crisis management exercises require careful consideration during their design due to their sociotechnical complexity. Likewise, their management and execution typically depend on a predefined and fixed schedule of events developed in advance. In the case of the RéSISTeCC exercise protocol it has been possible to establish a dynamic exercise management approach that adapts pedagogical objective and related injects to the participants’ actual responses and actions throughout the exercise. This paper provides an assessment of the activities carried out during the testing of the RéSISTeCC, describing the needs in cyber range topologies for IT services and how these infrastructures should be peopled by data according to the type of target organizations: local governments, small and medium-sized enterprises (SMEs), healthcare and long-term care facilities, fire and rescue services. We delve into the conception and progression of the exercise from its initial stages right through to crisis management and the importance of these different phases in the situation awareness, once the attack occurs. We explain real-time participants’ performance measurements and how they are used to contribute to the facilitators’ pool of injects based on pedagogical objectives and on the actual state of the IT infrastructure in relation to the cyberattack and the actions of those involved. These lessons learnt are based on the design, implementation, and testing of 28 exercises involving crisis response teams ranging in size from 2 to 30 people.

European cyber-resilience reporting requires product-security teams to form defensible reporting postures while vulnerability evidence remains incomplete, distributed and potentially contradictory. This paper presents a temporal vulnerability-evidence orchestration artefact that normalises software composition, supplier affectedness, exploitation intelligence, local applicability, deployment context and decision records into source-linked claims. The implemented model separates evidence-supported recommendation, deadline posture and the versioned EvidencePack/audit snapshot, while representing uncertainty, identity confidence and active conflict explicitly. Evaluation combines four controlled scenario families and matched controls, a 28-event manual-assisted PSIRT baseline, controlled sensitivity analysis, mutation testing, scale evaluation and a bounded public historical replay. Across the four primary scenarios, the orchestrated workflow populated the complete native 34-field EvidencePack schema (1.000 versus 0.824), although both workflows achieved complete common-field coverage and traceability. Both detected the sole seeded conflict, while conflict precision was 1.000 for the orchestrated workflow and 0.200 for the manual baseline. State-oracle conformance was 1.000 versus 0.258, and native EvidencePack generation was 1.000 versus 0.000, while a supplemental control confirmed that the manual workflow could produce an equivalent structured record bundle. The results support temporal orchestration as a means of improving decision continuity and audit reconstructability within the controlled study, without treating the resulting states as legal determinations or evidence of universal PSIRT superiority.

Adversary emulation is increasingly used to validate detection engineering, train SOC/CERT analysts, and assess defensive readiness against realistic attack behavior. However, reusable emulation artifacts remain difficult to produce and share: technique-level tests are often too isolated to capture attacker progression, while complete scenarios are frequently tied to specific cyber ranges, agents, or execution engines. This paper presents SCEN&R, an open, vendor-neutral, YAML-based specification for describing realistic adversary-emulation scenarios independently of their execution backend. SCEN&R models an emulation scenario as session-bound steps over reusable procedures, explicit variables, declared attacker infrastructure, and a directed acyclic execution graph with runtime conditions. This format integrates with MITRE ATT&CK and Atomic Red Team to reuse standardized technique-level procedures while extending them into coherent campaign-level scenarios. We describe the design of SCEN&R, its execution semantics, and a Python reference implementation that validates scenarios, resolves data flow, evaluates conditional branches, and supports their execution. Our goal is to provide a practical scenario language, together with a reference implementation, for portable adversary emulation across cyber ranges, purple-team exercises, detection validation, and security research.

Organizations run purple team exercises, find gaps, patch a few rules, and report success—but without a formal measurement framework, they cannot quantify how much detection improved or why specific techniques went undetected. Attack prediction models achieve strong F1 scores but produce no defensive output. CTI knowledge graphs map threats to countermeasures without ever testing whether those countermeasures fire on real telemetry. The gap between emulating an attack and demonstrating that detection improved remains open. We address this gap with a knowledge-graph-driven framework for instrumented purple teaming. A bidirectional knowledge graph (KG) encodes the full chain from ATT&CK technique to data source, detection strategy, analytic rule, log source, and sensor—not just the attack side. Each emulation step is annotated with expected telemetry and success criteria, so that when detection fails, backward traversal of the graph pinpoints why: a missing log source, an absent rule, or an inadequate configuration. After targeted remediation, the same scenario is replayed and improvement is quantified through operational key performance indicators (KPIs) anchored by detection coverage (𝐶). We evaluate on four threat profiles—nation-state espionage (APT29), financial cybercrime (FIN6), state destructive operations (Sandworm), and big-game-hunting ransomware (Wizard Spider)—all sourced from the public CTID adversary emulation library [1]. The primary endpoint is detection coverage (𝐶), evaluated with a fixed-threshold decision rule (Δ𝐶3 > 0.10 in ≥3/4 scenarios). The four-step detection maturity ladder raises strict coverage from 𝐶0 = 23.1% (out-of-the-box SIEM) to 𝐶3 = 71.6% (with KG-generated rules), with the KG-guided rule engineering step alone contributing Δ𝐶3 = +30.3 percentage points on average. The decision rule is satisfied in all four scenarios.

Modern product security teams must assess vulnerabilities across an expanding range of processor architectures— ARM, RISC-V, IBM s390x, PowerPC—yet most binary analysis toolchains target x86-64 exclusively. Binary translation offers a principled path: lift a foreign-architecture binary to LLVM intermediate representation, compile to x86-64, and apply existing analysis infrastructure without modification. A prerequisite for this approach is that translators produce valid, analysable x86-64 output at scale. We present the first large-scale empirical evaluation of this prerequisite, measuring three modern open-source translators (RetDec, Anvill, rev.ng) and the classical UQBT system against a corpus of 39,364 production ELF binaries drawn from ten Linux distribution releases across five architectures. Our results reveal a decisive architectural complementarity: RetDec achieves 79.1 % translation completion on ARM64 while failing on RISC-V (12.3 %) and s390x (3.2 %); Anvill reaches 94.1 % on RISC-V and 68.1 % on s390x while being largely ineffective on ARM; rev.ng provides the broadest coverage (85–100 % on ARM, 58 % on s390x) at the cost of a 36× speed penalty. A two-tool portfolio (RetDec + Anvill) achieves 55.9 % overall and 70.5 % on supported architectures at speeds compatible with PSIRT 24-hour deadlines under NIS2/CRA; adding rev.ng raises coverage to approximately 83–100 % on supported architectures (point estimate 93 %, 95 % CI ±9.8 % per architecture, derived from stratified sample of 𝑛 = 100). A complete PowerPC64 gap (0 %, all tools, 8,134 binaries) constitutes a significant risk for organisations operating IBM POWER infrastructure. Coverage figures are upper bounds on security analysis utility pending vulnerability preservation experiments on translated binaries. All data are publicly available on Zenodo [1].

With the growing development of information technologies, securing access to data has become a major concern. To address this issue, Data-Centric Security (DCS) aims to integrate an access control mechanism into the data itself, thereby adding a layer of security. However, addressing decentralization and dynamicity in architectures implementing DCS remains an open challenge. This paper proposes a structured review and highlights the limitations of existing approaches. Consequently, we propose a novel architecture of a decentralized and dynamic DCS. The proposed architecture relies on access control mechanisms, attribute-based encryption schemes, and trust-based mechanisms, enhanced with authentication, revocation, context-awareness, and decentralized delegation of access rights for subjects, formalized through six algorithms. The results highlight several strengths of the proposed D3CS architecture, as well as perspectives to be addressed in future work.

This paper proposes a structured methodology to transform real-world information disorder cases into formalized simulation scenarios. While existing frameworks provide taxonomies of tactics and interventions, they do not offer systematic procedures to derive executable models from empirical situations. We introduce a ten-step pipeline that extracts actors, narratives, assets, opinions, and systemic strains from real-world cases. The methodology is illustrated through a case study based on electoral information dynamics in Romania during November 2024. The resulting scenario demonstrates how qualitative observations can be translated into structured, simulation-ready models. To validate this approach, we implemented this scenario in Critical Masses, a software for simulating information disorder combat, that is based on Infodemiconium, a framework to model information disorder.

Zero Trust Architecture (ZTA) promotes access-control decisions grounded in identity, context, and least privilege rather than implicit trust based on network location. In hybrid environments combining on-premise and cloud resources, identity-driven micro-segmentation is often presented as a mechanism to reduce lateral movement, reachable surface, and post-compromise blast radius. However, security teams still lack controlled and reproducible comparisons that quantify what identity-aware, application-layer enforcement adds over traditional network-layer (L3/L4) ACL segmentation, and at what operational cost. This paper proposes a reproducible comparative framework and reports preliminary experimental results for a simplified hybrid information system. We compare two implemented protection variants: (i) traditional perimeter segmentation based on static L3/L4 network ACLs enforced by a firewall (pfSense), and (ii) identity-driven micro-segmentation enforced at the request level by an NGINX policy enforcement point coupled with an Open Policy Agent (OPA) decision engine, mutual TLS, and short-lived JSON Web Tokens. An optional third variant considers cryptographically attested workload identities using SPIFFE/SPIRE, in line with identity-tier guidance for cloud-native ZTA. The framework isolates the flow-control layer while keeping the same topology, application components, legitimate workflows, and offensive scenarios across variants. We operationalise security, complexity, and performance metrics, and run four offensive scenarios (unauthenticated access, database pivoting, cloud-side rebound, and application-layer credential abuse via SQL injection and brute forcing) under identical conditions. Preliminary results show that the network-ACL baseline blocks network-level pivoting yet lets an application-layer credential brute-force attack succeed, while identity-aware enforcement blocks all four offensive scenarios at a negligible median latency overhead (+4 ms, +0.07%). The objective is not to claim a complete Zero Trust deployment, but to provide a structured, reproducible basis for measuring when identity-aware enforcement adds containment value beyond static network zoning, and what policy-management and runtime costs it introduces.

Rule-based EDR detection struggles against APTs, whose signal lies in causal relationships between system entities rather than individual events. Provenance-graph GNNs address this, but existing systems are trained on research-grade syscall traces (e.g., DARPA TC) and have not been validated under production EDR constraints, which instrument a smaller subset of system interactions. We present a two-phase GNN pipeline for these constraints: Phase 1 pretrains a three-layer GAT on benign provenance graphs from an Elastic Endpoint sensor (≈2M nodes) using a masked graph autoencoder; Phase 2 fine-tunes a lightweight classification head on weak labels derived from MITRE Caldera operation reports, with the encoder frozen. On a self-collected dataset of four ATT&CK-mapped profiles (stealthy reconnaissance, exfiltration, cron and systemd persistence) and a held-out fifth, training on persistence and exfiltration enables detection of an unseen low-alert reconnaissance profile at AUPRC 0.635; within-profile evaluation reaches 0.648–0.973, against 4.1% alert precision and 11–25% recall on the evaluated profiles from the Elastic Security rule engine on identical telemetry.

Without coordinated intervention, AI systems will by default saturate most existing cognitive benchmarks by 2030, approaching or exceeding human expert performance across a broad range of measurable tasks. Using a joint hierarchical Bayesian model, we forecast frontier performance trajectories across 63 benchmarks spanning reasoning, mathematics, coding, scientific knowledge, and agentic capabilities. We find that nearly all current benchmarks (97% in our set, 80% CI: 95–100%) are on track to saturate within four years. This finding is robust to modeling choices: sensitivity analyses across 8 model variants and three saturation thresholds show that altering key assumptions yields qualitatively similar conclusions. Security-critical benchmarks (cybersecurity, autonomous AI R&D, biology, chemistry) show even faster trajectories, saturating before 2028. For the cybersecurity and cyberdefense community, these results raise two distinct categories of risk: (i) misuse of capabilities directly applicable to offensive operations (vulnerability discovery, exploitation, autonomous computer use, dual-use scientific knowledge), and (ii) loss of control arising from rapid progress on general capabilities (autonomy, agentic reasoning, and AI R&D automation), which together enable persistent autonomous agents able to accelerate their own development. While benchmark saturation does not equate to artificial general intelligence, these findings add to a converging body of evidence indicating that superhuman performance on cognitive tasks is approaching faster than commonly assumed. Acknowledging that we neither understand nor control the behavior of current GPAI models, this timeline leaves limited runway before global impacts. We outline implications for cyberdefense doctrine, international coordination, AI safety research, and evaluation practices.