Philipp Normann
PhD student in the Security and Artificial Intelligence Lab (SAIL) at TU Wien, advised by Daniel Arp
About
I work on machine learning for computer security. I study what AI systems can do in security tasks, and how we can make them more reliable and useful for defenders. A common theme in my work is verifiable feedback: results that can be checked without trusting the model. I use such feedback as a reward in training and as ground truth in evaluation. My current focus is on training and evaluating LLM agents.
I joined TU Wien in 2025, and my position is funded by the WWTF project BREADS. Before that, I worked for seven years as a data scientist at OTTO in Hamburg, where I built and maintained real-time recommender and fraud detection systems in production. I also co-authored three papers at ACM RecSys and ran the company's first Kaggle competition. I hold an MSc in IT Security and a BSc in Computer Science from FH Wedel.
News
- Dec 2026I will present two papers at ACSAC 2026 in Los Angeles: a workshop paper at WAITI on 7 December and the main conference paper on 10 December.
- Nov 2026Our ACSAC artifact received the badges Available, Functional and Reproduced.
- Oct 2026Our paper on IoT security guidelines was accepted at the WAITI workshop at ACSAC 2026.
- Sep 2026The full version of our work on local security agents was accepted at ACSAC 2026.
- Apr 2026Our paper on post-training for local security agents was accepted at the DeMeSSAI workshop at IEEE EuroS&P 2026 in Lisbon.
- Mar 2026Our paper on explaining security classifiers with sparse autoencoders was accepted at the XAISec workshop at IEEE EuroS&P 2026.
- Feb 2026Chasing Shadows: Pitfalls in LLM Security Research appeared at NDSS 2026.
Publications
Security and AI
-
Towards Reliable Local Security Agents: Verifiable Post-Training for Linux Privilege EscalationPhilipp Normann, Andreas Happe, Jürgen Cito, Daniel ArpPost-training a small open-weight model for Linux privilege escalation, with supervised fine-tuning and reinforcement learning on verifiable rewards. An earlier version was presented at the DeMeSSAI workshop at IEEE EuroS&P 2026.
-
From Exploits to Guidelines: What LLMs Can and Cannot Automate in IoT Threat IntelligencePhilipp Normann, Carlotta Tagliaro, Martina Lindorfer, Daniel ArpWAITI 2026 (workshop at ACSAC)Comparing IoT security guidelines with vulnerabilities exploited in practice, and testing which steps of this analysis an LLM can automate.
-
Chasing Shadows: Pitfalls in LLM Security ResearchJonathan Evertz, Niklas Risse, Nicolai Neuer, Andreas Müller, Philipp Normann, Gaetano Sapia, Srishti Gupta, David Pape, Soumya Shaw, Devansh Srivastav, Christian Wressnegger, Erwin Quiring, Thorsten Eisenhofer, Daniel Arp, Lea Schönherr
-
Towards Explaining Classification Models in Security with Sparse AutoencodersNils Ole Breuer, Lorenz Linhardt, Philipp Normann, Daniel ArpXAISec 2026 (workshop at IEEE EuroS&P) · Paper
Earlier work: recommender systems
-
Identifying Offline Metrics that Predict Online Impact: A Pragmatic Strategy for Real-World Recommender SystemsTimo Wilm, Philipp NormannACM RecSys 2025 · Paper
-
Pareto Front Approximation for Multi-Objective Session-Based Recommender SystemsTimo Wilm, Philipp Normann, Felix Stepprath
-
Scaling Session-Based Transformer Recommendations using Optimized Negative Sampling and Loss FunctionsTimo Wilm, Philipp Normann, Sophie Baumeister, Paul-Vincent Kobow
-
OTTO Recommender Systems Dataset: A real-world e-commerce dataset for session-based recommender systems researchPhilipp Normann, Sophie Baumeister, Timo WilmKaggle 2023 · used by more than 2,500 teams in the competition · Dataset · GitHub
Teaching
- Since 2025Machine Learning for Computer SecurityTeaching assistant at TU Wien, winter terms. I co-design and grade the exercises and help with the exams. In 2025, I gave a guest lecture on AI-based vulnerability discovery.
- 2026Attacking and Defending AI-based SystemsTeaching assistant at TU Wien, summer term. I designed challenges, partly with colleagues, and graded them.
Supervision
- 2026Diploma thesis, TU Wien. Co-advisor.
- 2026Design and Evaluation of a Local Open-Weight LLM-Based SSH Honeypot for Linux Shell EmulationBachelor thesis, TU Wien. Co-advisor.
- 2023Graph Neural Networks for Complementary Product Recommendations in E-CommerceMaster thesis at OTTO with FH Wedel. Co-advisor.Received the Hochschulpreis of the Wedeler Hochschulbund.
Talks
- Dec 2026Towards Reliable Local Security AgentsACSAC 2026, Los Angeles
- Dec 2026From Exploits to GuidelinesWAITI workshop at ACSAC 2026, Los Angeles
- Jul 2026Towards Reliable Local Security AgentsDeMeSSAI workshop at IEEE EuroS&P 2026, Lisbon
- Mar 2023Insights from the OTTO Kaggle CompetitionRecSys Meetup organized by NVIDIA
- 2022Deep Multimodal Recommendations: Solving the Cold Start ProblemMAIN 2.0 Conference
- 2018Real-Time Fraud Prognosis using Recurrent Neural Networkscode.talks · Recording
Service
Subreviewer: USENIX Security (2026, 2027), ACM CCS (2026), ACSAC (2026), IEEE SaTML (2026, 2027), RAID (2025), DIMVA (2026), ACM AISec (2025, 2026).
Outreach: Hamburg coordinator of meetup.ai, the largest AI meetup in Hamburg with more than 2,000 members, 2019–2025.