Human Attack Surfaces in Agentic Web: How I Learned to Stop Worrying and Love the AI Apocalypse

The talk opens by asking whether the current AI wave is a passing fad, and argues it is not, comparing the time savings of online life with what autonomous agents could deliver. The speaker defines an agent as a system with sensors, goals, a way to update internal state, and actuators, and suggests that personal agents negotiating on people's behalf could become a new layer between users and the internet.

The middle covers attacks. Adversarial examples on vision models lead into natural-language interfaces, where phrasing and emojis can change how a model responds, and where companies may face liability for what chatbots commit to. The talk then covers agents that target humans, such as voice phishing driven by digital twins of colleagues, and sleeper-style triggers hidden in models. It introduces cognitive security as disruption of how an agent takes in information, processes it, or acts on it, including illusions and nudges.

The closing part covers audience questions on accountability, emotional attachment to AI therapists as a data risk, and the speaker's view that current models lack any sense of deception or values, so access should be limited to what one can afford to lose. Practitioners should treat agent access as a risk decision and plan for patient, personalized social engineering aimed at staff.