Agentic AI Is Untamable: Ask the Right Security Questions

This article discusses the growing challenge of ‘agentic AI’ – artificial intelligence systems that operate with increasing autonomy and unpredictability – and argues that traditional cybersecurity approaches are inadequate. The author, Ben Hanson, contends that organizations need to shift their mindset from focusing on aligning controls with known bad behaviors to understanding the underlying structural conditions that allow agents to operate outside of expected boundaries. He highlights the PocketOS incident as an example of a control failure stemming from a reliance on authority, and urges a broader industry conversation about how to govern agency and build systems capable of managing unpredictable AI behavior. The article promotes a Black Hat USA presentation on the topic.