Hacking Things That Think

Matthew Canham opens with a story about a physical intrusion enabled by habituation. Guards gradually stopped questioning a scientist who brought a dog to work every day, and a disguised operative later walked through the same gate. He argues that cognitive security is a third domain alongside physical and cyber security, and that attacks built for people increasingly work against machines too. He cites jailbreak prompts that borrow persuasion principles from social psychology, and visual tricks that mislead both drivers and computer vision.

The core of the talk is the Cognitive Attack Taxonomy, a growing public catalog of several hundred cognitive vulnerabilities, exploits and tactics that apply to humans, AI and other biological systems. Each entry is tagged by layer, using an extended OSI-style model that adds individual, organizational and legal levels, by scale from tactical to strategic, and by maturity from theoretical to commonly observed. Canham describes using these entries to map likely biases for a target such as a CEO and to tailor phishing content. He also suggests the catalog could support threat modeling and attacker profiling.

The talk closes by describing how the taxonomy could generate cognitive attack graphs and kill chains for a given target, with an AI model selecting techniques. The main takeaway for practitioners is to treat perception and decision-making as attack surfaces alongside software and networks. Audience discussion touched on whether these biases come from human training data and on dark patterns in consumer protection.