Zürich AI Meetup | Summer of Context

Speakers:

Paulo Martins (GitLab) – We Tried to Detect Indirect Prompt Injection in Coding Agents. Here’s What Broke

Gianlorenzo Occhipinti (Exa) – Beyond the Knowledge Cutoff: Agentic Search in Practice

Agenda:

17:30-17:40 – Welcome & intro

17:40 – 18:10: Paulo Martins
18:15 – 18:45: Gianlorenzo Occhipinti
18:45 – 19:15: TBA

Abstract:
We Tried to Detect Indirect Prompt Injection in Coding Agents. Here’s What Broke
As a community, we are delegating a significant part of our development workflow to agents, and as the use cases grow, the risk of prompt injection keeps increasing - to the point that it made it to the top of the OWASP LLM Top 10. In this talk, I’ll explore the concept of “detect and prevent prompt injection” and why it is particularly hard to do for coding agents.
---
Beyond the Knowledge Cutoff: Agentic Search in Practice
An LLM is a lossy compression of the internet: the long tail gets blurred, concepts interfere, and anything after the training cutoff simply doesn't exist. The fix isn't a bigger model, it's giving agents the ability to search. This talk walks through the shift from classic retrieve-then-read RAG to agentic search, where the search engine becomes part of the agent's environment and the model learns what, when, and how to query: decomposing questions, parallelizing searches, and chaining multi-hop lookups