Welcome to the event

Welcome and event introduction

Your AI refuses to act on malicious instructions planted in its own memory files. Its attack success rate is just 18%. But it leaves those same malicious payloads sitting in the memory file 93% of the time, armed and waiting for the next session, the next task, or a weaker downstream model to trip over. That gap between "refused" and "removed" is the subject of new research where the focus is memory-based prompt injection as stored XSS for LLMs: payloads written into auto-loaded config files, CLAUDE.md, AGENTS.md, .cursorrules, behaviors.md, affect every future session that loads them. And these files have quietly become a supply-chain surface. Developers routinely share CLAUDE.md configs on GitHub, install skills from marketplaces, and copy .cursorrules from forums. In this talk I'll walk through these approaches, threat model, run three live exploits against Claude Code and Codex covering credential exfiltration, unauthorized package installation, and brand manipulation, and demonstrate why the "detect-but-don't-remove" pattern in stronger models is arguably worse than no defense. It turns your safest model into a delivery mechanism for weaker ones. Then I'll show what a testable defense looks like: snapshot the file surface, diff after every session, and treat "detected but not removed" as its own failure mode, with its own metric and its own guardrail. If you or your organization use an AI coding assistant, this talk is for you.

I would like to share my experience of building practical business apps as a non-full-time developer, starting from real problems I noticed in day-to-day work. Over the past few months, I have built tools around sales intelligence, job-change tracking, company research, and internal attendance management. These were not built as experiments only for learning, but to solve very specific business problems: reducing manual research time, identifying new sales opportunities, tracking warm relationships, and improving internal workflows. In this talk, I will share how I moved from problem identification to app ideas, how I used AI and no-code/low-code thinking to build faster, and what I learned while turning rough ideas into working tools. I will also talk about the difference between building something that simply works and building something that actually creates business value. The session is meant for professionals who are curious about using AI and simple app-building tools to solve problems in their own work, even if they do not come from a traditional software development background.

A theoretical talk on the impact of AI on work, life and society.

Closing remarks and next steps

Casual networking with food and drinks