ChiBrrCon 2027

Systems designed around AI and ML technologies are growing as fast in complexity as they are in popularity. These systems, especially in production and at scale, often follow a microservice architecture with LLMs, Rag databases, MCP Servers, clients, and more all communicating over the network. That doesn't even include all the other supporting technologies that are usually needed to orchestrate and deploy complicated systems. With all this complexity comes risk. However, discussions on the risks and vulnerabilities of these AI systems often leave out one of the most important perspectives; Traditional network design & security.  

During this session we will investigate the technologies and risks associated with running production AI systems by walking through several common architectures, from the simplest to among the most complicated. We will discuss vulnerabilities that are unique to these systems, such as prompt injection, as well as more traditional network and application security risks.  Also in focus is the attacker perspective. We will discuss how attackers view these systems, including their infrastructure targets, priorities, and motivations. Sprinkled throughout this conversation are real examples from penetration tests, red teams, and public disclosures.  This presentation might be a good fit for attendees interested in learning more about the network and systems architecture behind AI and ML systems, their security vulnerabilities and risks, or those who want to hear a few war stories on exploiting AI systems.