References:
1. A. Adamov and A. Carlsson, "Reinforcement Learning for Anti-Ransomware Testing," 2020 IEEE East-West Design & Test Symposium (EWDTS), 2020. Available: https://www.researchgate.net/publication/346942881_Reinforcement_Learning_for_Anti-Ransomware_Testing
2. A. Adamov, "Simulation Approach to Anti-Ransomware Testing," AMTSO Webinar, June 9, 2021. [Online]. Available: https://www.youtube.com/watch?v=-jtjCjc3r9I
3. "ChatGPT," Wikipedia, 2023. [Online]. Available: https://en.wikipedia.org/wiki/ChatGPT
4. OpenAI and Microsoft, "Disrupting Malicious Uses of AI by State-Affiliated Threat Actors," Feb. 14, 2024. [Online]. Available:
https://openai.com/index/disrupting-malicious-uses-of-ai-by-state-affiliated-threat-actors/
https://www.microsoft.com/en-us/security/blog/2024/02/14/staying-ahead-of-threat-actors-in-the-age-of-ai/
5. CERT-UA, "Кібератаки UAC-0001 на сектор безпеки та оборони із застосуванням програмного засобу LAMEHUG, що використовує LLM (велику мовну модель) (CERT-UA#16039)," July 10, 2025. [Online]. Available: https://cert.gov.ua/article/6284730

Outline:
1. Introduction
Motivation: Why AI changes the game for AV testing
Brief history of malware evasion techniques
Rise of adaptive, intelligent malware

1. Case Study: AI-Powered Ransomware with Reinforcement Learning
   Research background (EWDTS 2020, AMTSO 2021)
   How RL was used to train ransomware to bypass detection
   Testing methodology and key findings
   What this showed us about AV limitations

2. From Theoretical to Operational: The Rise of LLMs in Cyberattacks
   ChatGPT’s release and the shift in AI accessibility
   2024: Microsoft and OpenAI report state-aligned actors using LLMs
   Overview of threat actor usage: recon, scripting, social engineering

3. APT28's 2025 LAMEHUG Attack: A Real-World Turning Point
   Summary of CERT-UA’s discovery of LAMEHUG (July 2025)
   Analysis of how the malware generated system commands via prompts
   Challenges this poses to static, signature-, and behavior-based AV detection

4. Why Current AV Testing Fails
   Snapshot of AMTSO and other common AV test frameworks
   Inability to simulate adaptive AI-generated behaviors
   The “prompt-to-payload” gap in current test scenarios

5. Proposed Framework for Testing Against AI-Powered Malware
   Taxonomy of AI-powered malware (LLM vs RL vs Hybrid)
   Simulation architecture: AI malware agent => payload generator => command execution
   Benchmarking AV response to variability, prompt injection, and real-time adaptation
   Use of local vs public models in testing labs

6. Call to Action: Building AI-Aware Testing Labs
   Guidelines for security vendors and researchers
   How to create reproducible and ethical AI-malware testbeds
   Suggestions for future work (e.g., LLM adversarial red teaming)

Thesis / Takeaway:
Traditional antivirus testing approaches are insufficient to evaluate modern, AI-powered malware. This talk demonstrates why new, adaptive, AI-aware testing frameworks are urgently needed and proposes a concrete methodology for simulating and benchmarking AI-driven malware behaviors in a controlled lab environment.

Detailed Description:
The cybersecurity landscape is undergoing a radical transformation as LLMs and Reinforcement Learning (RL) techniques enter the offensive toolkit of advanced threat actors. While defensive security research has made significant progress in applying AI to detection and triage, adversaries are now turning the same technologies against us - introducing malware that can learn, adapt, and generate novel attack patterns on the fly.
This talk will present both foundational and newly emergent threats from AI-powered malware, beginning with early work (EWDTS 2020) in which reinforcement learning was used to teach a ransomware agent how to evade detection. We will then transition to real-world use cases, including the 2024 confirmation by Microsoft and OpenAI that nation-state actors were leveraging LLMs in cyber operations, and the groundbreaking APT28 campaign discovered by CERT-UA in July 2025. That attack operationalized a Hugging Face–hosted LLM (Qwen 2.5) to generate reconnaissance and document exfiltration commands dynamically, attempting to evade conventional AV defenses.
The second half of the talk will focus on testing methodologies. Today’s AV testing standards, from static signature tests to dynamic sandboxing, fail to account for the fluid, generative nature of AI-enhanced threats. We will expose these shortcomings and propose a new testing paradigm that integrates LLM-driven code generation, adaptive command execution, and adversarial prompt crafting.

Attendees will gain:
+A clear understanding of how AI-powered malware behaves differently from traditional samples
+A detailed proposal for how to construct meaningful, reproducible, and ethical AV tests for such malware
+Practical recommendations for AV vendors, red teams, and testing labs seeking to future-proof their detection capabilities

Intended Audience:
-Malware analysts, threat researchers, and reverse engineers
-AV developers and QA teams at endpoint security vendors
-Red/purple teams designing adversarial simulations
-Academics and practitioners exploring AI in offensive security
-Security product evaluators