Context, Retrieval & Compromised Agents

Two hands-on talks for people building with AI: one on getting the right context into your agents, and one on what happens when an agent gets compromised. Both include live demos. Building Better Context with Visual Intelligence & Semantic Retrieval *António Morais & Akbar Gumbira, Machine Learning Engineers, Elastic* Getting the right context to your agents helps get better results. Using smaller, more specialised models keeps the costs down. We will guide you through an example use case to transcribe, store and search thousands of scanned manuscripts. Then we will dig deeper to cover how we serve our in-house OCR, embeddings and reranker models, and how we designed our model serving service to work reliably for users across the world. António and Akbar build and operate the inference systems behind the Jina search foundation models in Elastic’s managed cloud. The Nine Lives of a Compromised AI Agent *Shade Obiti, Cybersecurity Analyst, Accenture* What happens when an AI agent gets hacked, and why isn’t revoking its access enough to stop it? Through a live demonstration of a real-world attack scenario, Shade shows how poisoned memory, shared identities and escalating access can let a compromised agent keep coming back, even after security measures have been put in place. You’ll see why traditional security controls aren’t always enough, and what developers and architects can do to build more secure agentic systems.