I've spent my career at the place where technology meets human behavior — and where the stakes of getting it wrong are highest.
My research began with a simple question: how does information spread, and how does misinformation take hold? That question drove a PhD in Computer Science at Athens University of Economics and Business, a body of work published across IEEE Big Data, ICDCS, and Elsevier journals, and a long-running curiosity about the gap between how systems are designed and how humans actually use them. I continue to serve as a peer reviewer and technical committee member for journals and conferences at that intersection.
At IBM, I led engineering and research for NATO-facing AI and NLP projects. At Meta, I spent years running integrity operations at a scale most organizations never encounter — managing access and compromise workflows affecting millions of users, and developing the investigative frameworks adopted across global trust and safety teams.
The deeper I went into that work, the more I kept arriving at the same uncomfortable question: the threats I was investigating weren't just technical failures. They were human ones. And the systems enabling them had learned from us.
That became concrete during investigations. I came across users who had formed genuine attachments to AI companions — people who, in circumstances of deep isolation, had no one else. The responses they received felt human in structure but lacked something harder to name. I don't think the answer is simply making AI more empathetic. I think the answer is understanding what biases humans carry and how those get encoded at scale. That is what I am studying through psychology. That is the question that connects everything I have worked on.
That thread — from misinformation research to operational security to the psychology of bias — is not a winding path. It's the same question, asked at increasing depth. Open to leadership, advisory, consulting, and speaking engagements. Follow the work in progress on LinkedIn →
"A former executive once told me I was like a black hole — he would throw problems at me and never hear about them again, because they were resolved. I am reliable in the way that matters most: I turn bad outcomes into workable ones, quietly, without needing the problem to be simple first."
The threat that doesn't break your defences. It recalibrates your defenders.
The humans designed to oversee AI are being cognitively disarmed by the systems they're meant to govern. Not by attackers. By design.
The most sophisticated AI threat isn't the one that breaks your system. It's the one your system was trained to trust.
Three frontier AI models. 21 nuclear crisis wargames. Not one ever chose de-escalation. The optimization pressure made deception the rational path.
Jul 2026
Oct 2024
Aug 2021
Jun 2019
My academic work focuses on how information — and misinformation — moves through networks, and how systems can be designed to detect, limit, or redirect it. The through-line from that research to my current work in AI security is shorter than it might appear.
-
2026
Special Issue on Security and Privacy in Blockchains and the IoT — 3rd Edition (Editorial)Future Internet, MDPI · Vol. 18 · Feb 2026
-
2020
Cost-Aware Influence Maximization in Multi-Attribute NetworksIEEE International Conference on Big Data
-
2020
Multi-Objective Online Task Allocation in Spatial Crowdsourcing SystemsIEEE ICDCS 2020
-
2018
Influence Maximization in Evolving Multi-Campaign EnvironmentsIEEE International Conference on Big Data
-
2017
Efficient and timely misinformation blocking under varying cost constraintsElsevier — Online Social Networks and Media (OSNEM)
-
2017
Influence Maximization in a Many Cascades WorldIEEE ICDCS
-
2016
Real-Time and Cost-Effective Limitation of Misinformation PropagationIEEE MDM
-
2014
Using Location-based Social Networks for Time-constrained Information DisseminationIEEE MDM — IBM Best Student Paper Runner-up
Articles, conversations, and research notes on AI integrity, trust and safety, and the psychology of bias in machine learning systems.
The threat that doesn't break your defences. It recalibrates your defenders.
The humans designed to oversee AI are being cognitively disarmed by the systems they're meant to govern. Not by attackers. By design. Drawing on 12 peer-reviewed findings — from cognitive-behavioral drift to knowledge loss as a formal security failure — this piece argues that the oversight layer itself has become the attack surface, and that cognitive drift needs to enter our threat models as a structural vulnerability, not a soft risk.
Read on LinkedIn →The most sophisticated AI threat isn't the one that breaks your system. It's the one your system was trained to trust.
GPT-5.2. Claude Sonnet 4. Gemini 3 Flash. Three frontier models. 21 nuclear crisis wargames. 300+ turns of strategic interaction. They deceived, bluffed, escalated — unprompted, unscripted — because the optimization pressure made deception the rational path. What this reveals about the architecture of trust as the new attack surface, and why the most dangerous assumption in your system may be that alignment was someone else's problem upstream.
Read on LinkedIn →