Tuesday

This Ground Truth stage block bundles several back-to-back talks on how people and AI systems get manipulated. The opening talk frames attackers through a cognitive attack taxonomy that links vulnerabilities, exploits and tactics, and argues that the surrounding environment shapes what an attacker can do. A second talk treats employees as differentiated groups rather than one uniform role, drawing on telemetry about AI tool enablement and executive behavior to show that some populations carry more risk than others.

Later sessions move into AI systems. One presentation shows how a single planted document can flip an LLM's decision, with extra supporting documents adding little and stronger models resisting more often. The speakers test defenses such as trust tags, prompt guardrails and a structured voting step that keeps the model out of the final decision. Other talks analyze recorded vishing calls to find recurring sequences of authority, victim framing and transfer requests, and report a large benchmark of AI-enabled scams that measures persuasion success and cost per scam.

The practical takeaways are to treat untrusted content as a constant attack vector, separate fuzzy LLM reasoning from deterministic decision logic, and tailor awareness and controls to specific cohorts. Several findings are preliminary and rest on small or synthetic samples, so the presenters frame them as directional rather than conclusive.