AI and cybersecurity

Game-theoretic defence

• Honeypots and deception: where to place decoys so that a rational attacker is most likely to reveal themselves (GameSec 2012; Advances in Information Security 2015). Recently: mapping the deception surface of MITRE ATT&CK, i.e. which attacker techniques decoys can and cannot cover (2026).
• Attack-graph games for optimal network hardening against strategic attackers (IJCAI 2015, AAMAS 2015, IEEE Intelligent Systems 2016, Computers & Security 2019).
• Intrusion detection against a rational adversary, a project funded by the Office of Naval Research Global where I was PI. Also: detecting data exfiltration and randomised operating points for adversarial classification.

Machine learning for detection

• Malware detection under concept drift: learning with predictions of future malware (2024).
• Robust malware classifiers trained with adversarial strings generated from perturbed latent representations (2021).
• Avast-CTU Public CAPE Dataset (2022), a public dataset of malware behaviour reports for research.
• Learning from raw structured data: hierarchical multiple-instance learning (Mill.jl, JsonGrinder.jl) and explaining such classifiers.

Agents and security

NASimEmu (ESORICS 2023 workshop) is a network attack simulator and emulator for training RL agents that generalise to network scenarios they have not seen. This is a step toward autonomous agents for penetration testing and defence. With the rise of LLM agents, the field is moving in two directions at once:

• agents for security, which triage alerts, investigate incidents and emulate attackers;
• security of agents, which must resist prompt injection, manipulation and poisoned memories.

Both directions call for the adversarial reasoning our group is known for.

Industry experience

I have worked on security problems with Cisco and Trend Micro, and inside Avast and Gen (Norton, Avast, AVG, Avira, LifeLock) as Principal AI Scientist and AI Architect (2019–2024). There I led the move of a malware detector used by 350 million people from hand-written rules to deep learning, and the AI architecture of an LLM-based scam analysis assistant. I also co-authored three US patents.