We will have keynote talks from the following world-leading researchers.

Keynote Talk I: Towards Quantum-safe privacy-preserving identification schemes

Abstract

In the past years, privacy-preserving digital identification schemes such as blind signatures and particularly anonymous credentials have seen rising interest both from industry and academia, due to their usefulness in online wallet, online age verification and similar applications. However, most currently suggested implementations are inherently built around assumptions that are broken by quantum computers. In this talk, I will take a look at the state of quantum-safe privacy preserving digital identification and the progress towards practicality that has been made in the past years.

Keynote Talk II: Beyond Prompt Inspection: Monitor What LLMs Do, Not Just What Prompts Say

Abstract

Large language models no longer just answer questions. They now drive agents and assistants that read untrusted web pages and documents, sustain long multi-turn sessions, and serve requests at industrial scale. Yet most defenses still inspect the prompt, scanning the input text for suspicious tokens or instructions. This talk argues that for deployed LLMs, the security-relevant event is often hard to identify at the prompt surface, and that we should instead watch what the model actually does. I will develop this idea through three projects. First, I will show that backdoor triggers need not be carried by visible input content. They can exploit overlooked input channels such as positional information, allowing normal conversations to activate hidden behavior. Second, I will discuss prompt injection in LLM agents, where the model may recognize a malicious instruction early but act on it only later. This mechanistic gap tells us where a detector should look. Finally, I will introduce a runtime monitoring approach that achieves high detection accuracy at almost zero added cost. Together, these projects suggest that securing deployed LLMs requires moving beyond input inspection. We should monitor how models represent context, where they make security-relevant decisions, and what behavior they reveal during serving.