May 14
Loading video player...
Agentic AI is changing how we build software, but it is also expanding the software supply chain in ways many teams underestimate. In this talk at KTH, I explore the security and maintenance risks behind modern agentic AI frameworks from a user perspective. The key point is simple: it does not matter whether code is written by a human or an AI agent. Once it enters production, someone must maintain it, secure it, and take responsibility for it. I discuss how agentic AI systems introduce new layers of complexity: model registries, memory stores, tool calling, runtime execution, connectors, observability systems, cloud services, and transitive dependencies. These layers make the trust boundary much larger than the application code itself. Using a real production Python project built with LangGraph and LangChain-related packages, I show how a small number of direct imports can pull in a much larger dependency graph. The result is a bigger blast radius, more integration surfaces, and new security concerns around prompt injection, state management, tool permissions, serialization, SQL injection, and runtime execution. The talk also covers why accountability still matters in AI systems. As IBM warned back in 1979, a computer can never be held accountable. Even in the age of AI agents, humans remain responsible for design decisions, deployment practices, dependency choices, and production safety. Key topics covered: • Agentic AI frameworks and their role in modern applications • Software supply chain risks in AI systems • LangGraph, LangChain, LangSmith, and AI orchestration • Transitive dependencies and dependency graphs • Prompt injection and tool-based attacks • Memory, state, and persistence risks in AI agents • Why agent autonomy increases the attack surface • Practical recommendations for securing agentic AI systems • Why teams still need dependency pinning, isolation, monitoring, and automation My main recommendation: treat agentic AI frameworks as production software infrastructure, not as magical wrappers around models. The magic is just code with better marketing. Connect with me: https://www.cesarsotovalero.net Subscribe to the channel: https://www.youtube.com/@cesarsotovalero?sub_confirmation=1 #agenticai #aiagents #aisecurity #SoftwareSupplyChain #secureai #SupplyChainSecurity, #LLMSecurity