Experience

Security engineering and research experience.

Agentic AI Systems Research

Zürich, Switzerland

Research

• Agentic AI systems: Technical experimentation and research on agentic systems, their architectures, and evaluation in real-world environments.
• Security and trustworthiness: Work across AI security, safety, reliability, and robustness, including the tools, permissions, context, memory, execution, and human-control boundaries of autonomous systems.
• Real-world deployment: Study of how autonomous systems plan, use tools, execute multi-step workflows, and can be evaluated for trustworthy deployment beyond model-only behavior.

Security Research Intern

Full-time·Düsseldorf, North Rhine-Westphalia, Germany · On-site

Internship

• Privacy-enhancing technologies: I researched data minimization, anonymous authentication, secure credentials, and privacy-preserving, policy-scoped authorization at Huawei Research, Data and Privacy Protection Lab.
• Policy-based anonymous tokens: For my Master’s thesis, I specified, implemented, and evaluated a protocol under a post-quantum security target. The goal is least-disclosure authorization, unlinkability, and local derivation of policy tokens, rather than cryptography in isolation.
• Application to Agentic AI workflows: I designed an authorization architecture in which trusted host components mediate model-proposed tool calls and select among anonymous, identity-bound, and step-up authorization for tools, APIs, and MCP servers. The architecture is analyzed, not deployed or benchmarked end to end.
• Measured trade-offs: At the evaluated endpoint, the policy-bound construction added 16.8% client redemption time including local derivation, 22.4% proof size, and 16.4% verifier time. These are artifact-level results rather than an end-to-end deployment benchmark.
• Operational cost model: In an illustrative same-primitive comparison, one tag-independent master credential became cheaper than repeated ordinary-token issuance from two policy-token uses onward, at 1,040.052 ms versus 1,264.948 ms for two uses.
• Research paper: I co-authored A Post-Quantum-Compatible Generic Construction of Policy-Based Anonymous Tokens, submitted to PKC 2027, covering correctness, freshness, unforgeability, and unlinkability under the stated building-block assumptions.

Undergraduate Researcher - Software Security

Benevento, Campania, Italy · On-site

research

• Conducted undergraduate cybersecurity research for my Bachelor’s thesis, "Extracting Equivocal Behaviours from Trusted Systems", focusing on software transparency, behavioral analysis, and software supply chain security.
• The main idea was simple: trusted software does not always behave exactly as users expect. Legitimate applications can perform actions that are also commonly associated with malware, without those actions necessarily being malicious. I investigated these ambiguous actions, known as Equivocal Software Behaviours (ESBs), and their possible impact on user security and privacy.
• Multi-sandbox pipeline: I developed a modular, asynchronous, multithreaded Python tool integrating Hybrid Analysis, VirusTotal, and ANY.RUN, then normalized their JSON reports across Windows 10 and 11 environments where supported.
• Behavioral taxonomy: I mapped observations from 60 reviewed MITRE ATT&CK techniques into a 12-category taxonomy of Equivocal Software Behaviours.
• Study scope and result: I analyzed 36 widely downloaded goodware binaries across ten software categories. System and resource discovery appeared in 100% of the analyzed systems, showing why vendor reputation alone is insufficient evidence of safe runtime behavior.
• Reusable analysis: I produced structured JSON data and Pandas/Seaborn visualizations to make behavior comparable across sandbox outputs.