# Renato Mignone: professional background

Canonical: https://renatomignone.github.io/profile/
Editorial review: 2026-10-30

## Agentic AI security

I research how AI agents fail and how to make them secure and robust. My work draws on systems security, threat detection, and privacy-preserving authorization.

## How I came to security

My interest in security began during my Computer Engineering degree at the University of Sannio, where I built a Python pipeline to study trusted software through sandbox reports. That research led me to pursue an **MSc in Cybersecurity Engineering at Politecnico di Torino**, which I completed on October 2, 2026 with **110/110 cum laude**.

The degree gave me practical experience across systems security, secure web applications, networks, and machine learning for cybersecurity. The AI and Cybersecurity course helped shape my interest in how AI can support security analysis and how AI systems themselves can fail.

At **Huawei Research**, I worked on privacy-preserving authorization for my master’s thesis. Alongside that research, I built open-source tools for AI agents, including **VitaeContext**. I now conduct security research with the **Agentic Systems Lab at ETH Zurich**, studying how manipulated observations can affect adaptive agents’ later decisions and evidence acquisition. I want to keep developing these interests through security engineering and research.

## Learning to investigate software behavior

During my bachelor's degree in Computer Engineering at the University of Sannio, I worked with the university's cybersecurity research group. I wanted to understand what trusted applications actually do when they run. Legitimate software can perform actions also associated with malware, and a trusted reputation alone gives an incomplete picture of its behavior.

For my thesis, I built a Python pipeline to collect and normalize sandbox reports from Hybrid Analysis, VirusTotal and ANY.RUN. The work combined asynchronous execution, multithreading, structured data and behavioral analysis. I reviewed 60 MITRE ATT&CK techniques, organized observations into 12 categories, and studied 36 binaries across ten software categories. The conclusions are bounded by that sample and the sandbox environments. The [thesis and dataset](https://doi.org/10.5281/zenodo.19262869) are public.

That experience made me want to study cybersecurity further. It also gave me a way of approaching security problems: look at observed behavior, connect it to the underlying system, and state what the evidence supports.

## Building broader engineering foundations

I moved to Turin for the MSc in Cybersecurity Engineering at Politecnico di Torino. The English-taught programme covered systems, networks, applications, hardware, verification, cryptography, forensics and AI. I graduated on October 2, 2026 with 110/110 cum laude.

The projects let me work at several levels of a system. In [Linux eBPF research](/projects/Linux-eBPF-Verifier-Bypass-Research/), I traced cases from patches and verifier decisions through bytecode and runtime behavior. The analysis found a taint-tracking bypass and a bounded intra-packet overflow, without establishing privilege escalation. In the embedded project, I implemented a QEMU machine model for an NXP board and ported a FreeRTOS workload to exercise interrupts and peripherals.

Application and network projects added different trust boundaries. I built a forum with server-enforced ownership checks and TOTP-gated moderation, and compared VPN topologies using packet captures. These were academic projects, each with a specific implementation and evaluation scope.

## Connecting AI and cybersecurity

The AI and Cybersecurity course was an important part of how my interests developed. I explored how models can support security analysis through [threat-detection experiments](/projects/AI-Driven-Threat-Detection-Research/) and analysis of SSH honeypot sessions.

I trained and compared neural, graph, sequence, anomaly-detection and transformer models. The evaluations included imbalanced classes and shortcut learning. In one network-flow experiment, changing a destination-port pattern caused Brute Force recall to fall from 94% to 5%. That result made the limits of aggregate performance particularly concrete: a detector can appear strong while relying on a brittle feature.

I also built ARIA, a digital-forensics training game where players check AI claims against evidence. It reflects a related concern: AI assistance is useful only when the person or system using it can verify what supports an answer.

## Privacy-preserving authorization at Huawei Research

From February to August 2026, I worked as a Security Research Intern at Huawei Research's Data and Privacy Protection Lab in Düsseldorf. My master's thesis concerned policy-based anonymous tokens: how to prove scoped authorization facts while limiting unnecessary identity disclosure and persistent tracking.

I specified, implemented and evaluated a protocol under a post-quantum security target, and co-authored a research manuscript on privacy-preserving authorization. Cryptography was a mechanism for achieving privacy and authorization properties within this work.

I also designed and analyzed a proposed authorization architecture for agentic workflows. Trusted host components would mediate model-proposed tool calls and select appropriate authorization. This architecture remained a proposal; I do not present it as an implemented MCP gateway or an end-to-end deployed system.

## Building tools and working with teams

Alongside research, I build open-source software. With Elia Innocenti, I co-created [VitaeContext](/projects/VitaeContext/), a local-first career-context framework. It exposes focused context through agent skills, a CLI and a stateless MCP server. The work gave me practical experience with context routing, structured memory, JSON-RPC interfaces and provider portability. Its purpose is career-context management; the project is not a cybersecurity product.

At Hack Apertus 2026, our four-person team built [B2Bern](/projects/B-to-Bern/), a research navigator connecting company needs to public expertise in Bern. We combined an Apertus-based agent workflow, hybrid retrieval, evidence assessment and cited reporting. The team won the Bern Economic Development Agency challenge and placed second overall. It was a hackathon prototype, with shared team ownership.

I also contributed to Budget HQ, the first-place IEEE-HKN international hackathon project, and supported technical operations and partnerships for the Mu Nu chapter from April 2025 to June 2026. I became a lifetime IEEE-HKN member in December 2025. I am a community member of OWASP's GenAI Security Project, following its agentic-security work.

## Current research and next steps

Since September 2026, I have been conducting security research with the Agentic Systems Lab at ETH Zurich. My work investigates how manipulated observations can influence an adaptive agent's later decisions and the evidence it chooses to acquire. I develop threat models and experimental methodology for studying failure propagation and recovery, connecting individual observations to the behavior of the wider agent system.

I am also developing [From LLMs to Secure Agents](https://renatomignone.github.io/From-LLMs-to-Secure-Agents/), an educational guide that explains agent systems before revisiting their components through threats and controls. I designed a two-pass roadmap and built tools for scoped authoring, source records, validation and publishing. The guide remains in development.

My focus is AI security: investigating failure modes in agentic systems and evaluating how to make them secure and robust. I am interested in research and engineering roles working on agent behavior, tool access, context, permissions and security controls. My background in systems security, secure applications, threat detection and authorization supports this direction. I am open to relocation and opportunities to develop these skills through practical engineering and research.

You can find my code on [GitHub](https://github.com/RenatoMignone), connect on [LinkedIn](https://www.linkedin.com/in/renato-mignone/), or [email me](mailto:renato.mignone@gmail.com).
