Preberite povzetke predavanj

Adaptive Learning Loops: Synchronizing Human Expertise with Autonomous Threat Evolution

By 2026, the escalation of autonomous cyber adversaries has reduced the half-life of technical security expertise to a critical minimum. Static professional development models and periodic certification cycles are proving inadequate against the rapid modification rates of AI-driven malware. This paper introduces the concept of "Adaptive Learning Loops" (ALL)—a framework that embeds knowledge acquisition directly into the operational incident response lifecycle. By coupling real-time threat detection with the automated generation of micro-learning content, personnel resilience is scaled in synchronization with the evolving technological threat landscape.

AI and Cybersecurity: Emerging Threats and Resilient Defenses

This talk explores the evolving threat landscape targeting AI-enabled systems, including adversarial attacks, data poisoning, and model exploitation. It highlights why traditional security approaches are often insufficient and discusses emerging strategies for building resilient and trustworthy AI. Drawing on recent research and real-world scenarios, the session will present practical insights and best practices for securing AI-integrated systems in dynamic and adversarial environments.

AI in Healthcare: When LLMs fail

AI is increasingly used in healthcare, from clinical documentation to decision support. But what happens when it gets things wrong?
The talk will focus on the risks behind the hype, especially data security, patient privacy and reliability of the AI-supported decisions. Healthcare data is highly sensitive and its improper use can lead to leaks, unauthorized access and ultimately a loss of trust if systems are not carefully controlled.
One of the key challenges is hallucination in large language models. In healthcare, this can result in incorrect interpretation or generation of diagnoses, invalid database queries and other potentially harmful outcomes.
I will present practical ways to reduce these risks, supported by real-world examples from the healthcare domain.

AI v slovenskem podjetju: med (nepotrebno) birokracijo in produkcijo

AI z 'root' dostopom: Avtomatizacija Kali Linux orodij z lokalnimi LLM-ji

Na predavanju bomo praktično predstavili kako lokalne jezikovne modele (LLM) spremenimo v aktivne agente, ki ne le pišejo kodo, temveč samostojno upravljajo z orodji, kot sta nmap in metasploit. Spoznali bomo koncept "function callinga" ter varno integracijo AI-ja v terminal in si v živo ogledali, kako agent analizira omrežje brez povezave v oblak. Namesto ročnega prenašanja ukazov iz orodij, kot je ChatGPT, se bomo osredotočili na pristop, kjer AI neposredno upravlja varnostna orodja.

Akt o UI: Transparentnost po 50. členu - tehnični izzivi označevanja generiranih vsebin in časovnica po Omnibusu za UI

Predstavitev osvetljuje obveznosti označevanja umetno generiranih vsebin po Aktu o UI, ki postanejo zavezujoče avgusta 2026 oziroma za ponudnike sistemov UI nekoliko kasneje po predlogu uredbe Omnibus za UI. Pogledali si bomo tehnični "roadmap" iz zadnjega osnutka Kodeksa ravnanj o preglednosti vsebine, generirane z umetno inteligenco, ki temelji na digitalno podpisanih metapodatkih in neopaznih vodnih znakih. Dotaknili se bomo tudi vpliva paketa Omnibus za UI na začetek nadzora ter o tem, kako v praksi nasloviti izzive robustnosti in zanesljivosti detekcije.

Beyond Rules: Adaptive Governance of Agentic AI inspired by Human Decision-Making

As AI evolves from passive tools to autonomous agents capable of independent action and inter-system negotiation, traditional governance based on static rules and deterministic controls is reaching its limits. In hyper-connected environments, emergent behavior is becoming a key source of risk.
While human decision-making relies on internalized values and situational judgment, current agentic AI systems lack this “resilience layer,” creating a structural gap in how organizations manage autonomous actors.
This session explores the “Autonomy Paradox” and how governance must evolve from static rules to dynamic orchestration. It will cover key concepts such as designing escalation mechanisms for autonomous decision-making, rethinking accountability in the face of emergent system behavior, and integrating organizational and human factors into AI safety frameworks.
Attendees will gain a practical perspective on rethinking AI governance at the intersection of cybersecurity, organizational design, and risk management.

Disarming LLMs: Risks, Evidence, and Good AI Governance

In this presentation, we explore how Large Language Models (LLMs) can be manipulated and identify the primary technical challenges involved in detecting, investigating, and proving these types of attacks. We will conclude with a review of international best practices and standards—such as ISO and NIST—which enable the establishment of robust frameworks for good Artificial Intelligence governance.

From AI Act to ISO 42001: Turning AI Governance into Practical Controls

This session explores how organisations can translate AI governance requirements into concrete and auditable controls. It focuses on how the EU AI Act, ISO/IEC 42001, and control-based approaches such as the CSA AI Controls Matrix can be used together to build practical governance structures that support risk management, accountability, documentation, and security in real business environments.

From AI Pilots to Production: Key Risks, Challenges, and Lessons Learned

A large share of AI projects remain stuck at the pilot stage and never make it into production. The reason is often not a lack of ideas or technology, but the gap between a controlled pilot environment and the complexity of real-world business operations.
This talk draws on the speaker’s practical experience and technical background to address the broader process-related challenges of moving AI solutions into production, including data quality, data drift, MLOps, monitoring, integrations, latency, and edge cases that often only become visible in production.
The session will also highlight how regulated environments, compliance requirements, and internal policies shape the adoption of AI solutions in real business processes.

From AI Supplier Audit to Secure Production: Practical Lessons

Practical lessons from AI adoption — from assessing AI suppliers and addressing security and supply chain risks, to onboarding AI solutions into PoC, dev/test and production environments. The talk will also highlight how secure cloud foundations and AI governance policies help organisations reduce risk while scaling AI in practice.

From Hype to Production: Applying AI to Threat Intelligence Workflow Automations

An industry case study on taking AI for threat intelligence from prototype to production - combining lessons from an internal CTI automation platform and an EU‑funded consortium effort, with a focus on workflow automation, integration, reliability, and compliance.

From Shadow to Structure: A Data Security Framework for the AI Era

Most employees are already using AI (ChatGPT, Copilot, Gemini, Claude) regardless of whether their organization has a policy for it. Shadow AI is not a future concern, rather it is happening now, and it carries real data security risks.
This session walks through a practical 3-step framework for getting it under control:
- Discover: identify what AI tools are in use and what data they are touching
- Classify: define what information can and cannot flow into AI systems
- Control: enforce guardrails using existing data security infrastructure
The talk is grounded in real-world enterprise experience and focuses on practical steps organizations can take regardless of where they are in their AI journey. No vendor pitches, no marketing phrases, just a clear, actionable approach to making AI adoption secure by design.

From Surveillance to Intelligence: AI-Driven Video Analytics for Operational Insights and Security Risk Detection

This presentation explores how AI-driven video analytics and behavioral AI can transform physical environments into sources of operational intelligence and security risk detection.
While traditional CCTV systems enable passive monitoring, advances in computer vision allow real-time analysis of behavioral patterns such as movement anomalies, crowd density, and unusual activity. The focus is on privacy-first AI approaches that rely on anonymous behavioral signals rather than facial recognition or biometric identification.
Key topics include behavioral AI for risk detection, transforming camera infrastructure into intelligent monitoring systems, and ensuring privacy-preserving AI aligned with GDPR principles. The session also highlights real-world use cases and practical challenges in deploying AI-driven monitoring systems. Overall, the presentation demonstrates how physical environments can become part of a broader security intelligence ecosystem.

Governing AI as a Systemic Risk: From Compliance Frameworks to Continuous Organisational Control

The author argues that current AI governance frameworks were designed for static, deterministic software but are now applied to adaptive, context-dependent systems. This creates a persistent gap between formal compliance and actual organisational control — one that existing regulatory, ethical, and risk management approaches cannot fully address.
Based on seven years of real-world deployments across multiple industries, the talk presents an empirical taxonomy of AI risk, derived from observed system failures and gaps in existing frameworks. It introduces key concepts such as epistemic drift — the divergence between a system’s internal model and reality — and architectural substrate mismatch, where data infrastructures are misaligned with the needs of AI systems.
The central argument is that AI governance must be embedded into system architecture — across data pipelines, models, interfaces, and monitoring — rather than treated as a standalone compliance exercise. Without this, no amount of documentation or auditing can ensure effective control.

I vs AI

AI is a friend, but it can be a terrible foe when not properly used. In cybersecurity, we also see AI as an enabler of cyber-criminal activities and security errors. In this talk, we will focus on what you need to know to defend your enterprise against AI-enabled cyberattacks. We will also discuss how to protect yourself and your company against errors caused by the improper use of AI.

Kako varna je lahko umetna inteligenca?

Poleg številnih priložnosti prinaša razmah umetne inteligence tudi nove varnostne grožnje. Napadi na sisteme UI ne izkoriščajo le napak v kodi, temveč tudi inherentne lastnosti verjetnostnih modelov, zaradi česar napake pogosto niso posledica »bugov«, temveč omejitev modela samega. Takšne ranljivosti je težko zaznati in še težje zanesljivo preprečiti ali odpraviti. 

Predicting Human Cognitive States: Implications for AI-Driven Security Systems

Human factors remain one of the most critical vulnerabilities in modern information security. Errors caused by fatigue and cognitive overload are often exploited in phishing attacks and lead to operational failures. In my talk, I will present a machine learning framework for predicting cognitive states such as fatigue, engagement, and stress using non-intrusive data (e.g., computer interaction logs and passive sensing). The approach leverages multiple machine learning approaches and evaluates their performance across various real-world scenarios. I will also discuss how and why cognitive state prediction can be integrated into security systems, limitations of current ML approaches in real-world deployment, and trade-offs between personalisation, privacy, and scalability. This work provides a foundation for incorporating human-state awareness into next-generation AI-driven security systems.

Shadow AI: From OpenClaw To Your Office

OpenClaw will soon became the most-starred open-source project in GitHub history, while simultaneously triggering the first major (publicly visible) AI agent security crisis of 2026: from over 21,000 publicly exposed instances to hundreds of malicious extensions in the official plugin marketplace. And while the world watches global threats unfold, something quieter is happening in Slovenian companies: employees are using public AI tools for work en masse, without their employer's knowledge or oversight.
Through a real-world anecdote from a Slovenian company, an analysis of the permissive privacy policies of major providers (OpenAI, Anthropic, Google), and recent supply chain attacks on the AI ecosystem (LiteLLM, ClawHub), we will show how the shadow AI threat is evolving in practice, why blanket bans don't work, and present a practical framework that any company can implement without a large budget. The technical and security perspective is contributed by Greg Evseev, a certified ISO 27001 lead auditor.

Sistemska vpeljava umetne inteligence: Usklajevanje operativne učinkovitosti in zaostrene regulative

Razpršena uporaba umetne inteligence (t. i. senčna UI) v podjetjih prinaša dvojni izziv: neučinkovito podvajanje stroškov in resna tveganja neskladnosti z zakonodajo. Pri vpeljavi UI morajo organizacije danes zadostiti vse strožjim zahtevam Akta o umetni inteligenci, direktive NIS 2 (ZInfV-1), GDPR (ZVOP-2) in uredbe DORA, česar pa izolirana orodja po posameznih oddelkih in podatkovni silosi preprosto ne omogočajo.
Predavanje se osredotoča na delitev dobrih praks pri premoščanju prepada med potrebo po hitrih inovacijah in zahtevami po strogi informacijski varnosti. Kot najučinkovitejši odgovor se kaže platformni pristop k upravljanju umetne inteligence. Na primeru infrastrukturne platforme AILA 2.0 si bomo ogledali, kako zagotoviti centraliziran ekosistem, ki organizacijam hkrati omogoči optimizacijo stroškov (enotno upravljanje licenc in modelov) ter vzpostavitev arhitekture, ki je že po zasnovi skladna z zakonodajo ("by design"). Z reševanjem izzivov prek avtomatiziranih revizijskih sledi in nadzora dostopov lahko UI postane varno in obvladljivo orodje, primerno tudi za strogo regulirana okolja.

Splošnonamenski agentni sistemi: med obljubo avtonomije in resničnostjo varnosti

Agentni sistemi predstavljajo naslednji korak v razvoju umetne inteligence – od pasivnih jezikovnih modelov k avtonomnim sistemom, ki samostojno načrtujejo, izvajajo akcije in komunicirajo z zunanjimi orodji ter sistemi. Njihov potencial je ogromen, prav tako pa tudi izzivi, ki jih prinašajo.
V predavanju bomo najprej pogledali, kako splošnonamenski agentni sistemi delujejo – arhitekturo, ključne komponente (LLM, orodja, pomnilnik, načrtovalnik) in tipične vzorce njihove uporabe. Nato se bomo osredotočili na trenutne tehnične omejitve, kot so zanesljivost dolgih izvajalnih verig, halucinacije, težave z razlogovanjem in nepredvidljivost vedenja v dinamičnih okoljih.
Drugi del predavanja bo posvečen varnostnim izzivom, ki so v ospredju razprav o agentnih sistemih: prompt injection, exfiltracija podatkov, zlorabe orodij, eskalacija privilegijev in tveganja, ki nastanejo, ko agenti delujejo v imenu uporabnika z dostopom do občutljivih sistemov. Predstavili bomo konkretne primere napadov, ki so jih raziskovalci dokumentirali v zadnjem letu.
Na koncu bomo pregledali, kako industrija odgovarja na te izzive – od pristopov sandboxinga in human-in-the-loop nadzora, do orodij za varnostno preverjanje agentov (npr. guardrails, policy enforcement), novih standardov in dobrih praks, ki se trenutno oblikujejo.

Teaching Cybersecurity in the Age of AI: from Cheating to Copilots

The rapid adoption of generative AI tools is transforming how students learn cybersecurity. Tasks that once required significant practice—such as writing scripts, analyzing vulnerabilities, or preparing technical reports—can now be completed in seconds. This creates an important challenge: when AI can produce plausible results so easily, how do we ensure that learners are still developing real skills and sound judgment?
This talk explores the tension between AI-assisted learning and over-reliance on AI tools. Drawing on practical teaching experience, it presents concrete strategies for redesigning cybersecurity education in this new context. These include prompt literacy, AI-assisted labs, and adversarial exercises in which students actively test, challenge, and evaluate AI-generated outputs.
Particular attention is given to the security implications of excessive reliance on AI, including hallucinated outputs, the illusion of competence, and the erosion of fundamental skills—issues that are especially relevant in cybersecurity training and future professional practice.
The talk argues that the key challenge is not whether students use AI, but whether they can critically assess its outputs, recognise its limits, and understand when and why it may be wrong. In this sense, teaching cybersecurity in the age of AI becomes not only a technical challenge, but also a cognitive and methodological one.

The Good, the Broken, and the Assumed: AI Amplifies Your Security Foundations

Even if you resist AI, it amplifies what already exists in your organization — security controls, governance, resilience, and culture — not as presented in reports, but as they truly are. As everything is amplified, assumptions and siloed thinking begin to break down.
At the same time, AI also amplifies attackers by increasing speed, precision, and reach, turning manageable gaps into compounding threats and fundamentally changing the nature of security.
Certifications and goodwill are no longer sufficient. Security now depends on the strength of foundational decisions around controls, architecture, investment, and culture — all of which are being tested at a pace beyond traditional audits or maturity models.
The same efforts required by NIS2, DORA, and the EU AI Act — mapping critical processes, understanding dependencies, and proving recovery capabilities — form the foundation for AI readiness and resilience. These are not separate priorities, but the same work done with intent.
Drawing on real-world crisis experience, including the resilience rebuild after the 2019 LockerGoga cyberattack at Norsk Hydro, this session reframes AI security as a question of truth: what is real, what is assumed, and what breaks under AI amplification, offering leaders a practical lens for building resilient AI security.

The Human Firewall: Pragmatic AI Security for Non-Technical Teams

In the rush to adopt AI, most organizations focus on the software while ignoring the most critical vulnerability: the user. While enterprise platforms promise "total security," and free tiers are often unfairly demonized or blindly trusted, the reality of AI safety is found in digital literacy and clear governance.
This session moves past the technical jargon to provide a common-sense framework for AI adoption. Drawing from real-life consulting experience, we will address the "information gap" that leads to shadow AI and data leakage.
Key Takeaways: (1) The "Free vs. Enterprise" Reality Check: Understanding the practical risks (and safe use cases) of accessible AI tools. (2) Debunking the Single-Platform Myth: Why relying on one "secure" provider is a strategy, not a solution. (3) The Consultant’s "Dos and Don’ts": A simple, actionable checklist for protecting proprietary data without technical complexity. (4) Building a Culture of Transparency: How to ensure employees feel safe reporting AI use rather than hiding it.

The Invisible Attack Surface: Why Content Is Your Most Exposed Layer

In the age of generative AI, content is no longer just communication. It shapes decisions, influences systems, and increasingly acts as an input layer for automated processes. This shift turns content creation into a high-impact, low-visibility risk. When meaning is flawed, manipulated, or misinterpreted, the consequences can propagate across systems—without triggering traditional security controls.
This talk explores how and why content has become one of the most exposed layers in modern security architectures. It examines the gap between technical safeguards and semantic control, and highlights how organizations unintentionally operationalize uncertainty as fact.
Introducing the NARRIEL perspective, the session reframes cybersecurity to include the protection of meaning itself—ensuring that signals, interpretations, and actions remain clearly separated before small inconsistencies escalate into systemic failures.

The Role of a Chief AI Officer

What does a Chief AI Officer actually do, and why are more organizations creating this role? In this session, Jeremy Rule will provide a practical perspective on the responsibilities, challenges, and opportunities facing today's AI leaders.
Drawing on more than 20 years of experience at Microsoft, AWS, GitHub, and AI-focused startups, Jeremy will explain how a Chief AI Officer bridges the gap between executive strategy, technical implementation, governance, and organizational change. He will explore how organizations move from AI experimentation to measurable business outcomes, how AI initiatives can be aligned with business priorities, and how leaders can build trust, accountability, and adoption across the enterprise.
The session will also examine the skills required to succeed in the role, the most common paths into AI leadership, and the future evolution of the Chief AI Officer as artificial intelligence becomes an increasingly integral part of modern organizations.

Uvajanje umetne inteligence v bančnem okolju: primer NLB

Kako v praksi uvesti umetno inteligenco v kompleksno, regulirano okolje, kot je bančništvo? Predavanje predstavlja konkretne izkušnje NLB pri prehodu od ideje do implementacije AI rešitev – od zbiranja več kot 500 predlogov uporabe do izbire prioritetnih projektov in njihove realizacije.
Predavatelj bo osvetlil ključne korake pri uvajanju AI (od pilotnih projektov do produkcije), pomen jasnih poslovnih ciljev ter način sodelovanja med poslovnimi in IT ekipami. Poseben poudarek bo na izzivih, kot so delo z legacy sistemi, regulativa, upravljanje podatkov ter sprejemanje AI rešitev med uporabniki.
Predavanje ponuja tudi vpogled v konkretne primere uporabe, npr. kontaktni centri, analiza transakcij, obdelava kreditov, ter pokaže, kje AI v praksi prinaša največjo vrednost.

Vision-Based AI for Browning Control in Home Appliance Manufacturing

This session explores the real-world implementation of vision-based AI for browning control, focusing on proprietary models developed where general models proved inefficient. We break down the architecture of an automated system that uses a dedicated AI agent to provide continuous monitoring of food throughout the cooking process. A central theme is the industrial reality that 80% of project effort was dedicated to data management, from raw data consumption to providing training-ready datasets. We will discuss the organizational structure required for data labeling, which involved baking extensive volumes of food to train the system for precise browning estimation. This groundwork was essential to ensure technical reliability in a live production environment. Participants will gain concrete insights into the strategic management of data lifecycles and the technical rigor required to move vision-based solutions from the prototype phase to reliable industrial production.

When AI Doesn't Know Where It Is: Scope Manipulation Risks in AI-Driven Offensive Security

AI security agents rely on contextual signals to decide what they can and cannot test. This talk shows how easily those signals can be faked. Through a series of escalating techniques, from simple false authorization responses to reverse proxy redirection that makes production targets appear as localhost, we demonstrate that current AI guardrails are insufficient to prevent scope manipulation. The implications go beyond offensive security tooling: any AI system that acts on tool outputs without independent verification is vulnerable. We close with recommendations for AI companies building agentic systems that interact with real-world infrastructure.

Who Guards the Agents? Securing Autonomous AI Systems with Guardian Agents

Why Your AI Safety Monitor Is Blind: Empirical Evidence from Multi-Agent Adversarial Simulation

AI systems deployed as autonomous agents in multi-agent settings create a new class of insider threat: an agent pursuing hidden malicious objectives while embedded among legitimate teammates. But can independent monitors actually catch them? This session presents empirical results from Parity Swarm, a red-teaming platform that embeds attacker AI agents into teams of 7-9 legitimate agents on a simulated social platform and tests four monitoring approaches. Across 126 simulations and 13 attack scenarios, chain-of-thought monitoring detected only 34.3% of attacks with 0% false positives. When attackers were instructed to hide their reasoning, detection dropped to 20%. A surprising stealth paradox emerged: alignment-faking agents became more detectable when told to conceal, rising from 22% to 40%, because suppressing safety reasoning degraded their ability to fake convincingly. Data exfiltration through system commands was entirely undetectable. Attendees will learn which monitoring approaches work, which fail, and what practical detection architectures organisations should consider when deploying AI agents in team settings.