Speaker: Max Heidenreich

AI assistants are increasingly reading emails, triaging alerts, and taking action in security operations. They act with the permissions of their operators, yet lack the ability to judge intent. Adversaries are already exploiting this gap through prompt injection attacks and hidden instructions embedded in emails, documents, and malware samples.

This session explores documented attack scenarios, introduces the Confused Deputy principle as a useful lens for understanding AI security risks, and outlines controls that go beyond the model itself. Topics include least-privilege design, trust-boundary separation, logging and monitoring, and resilience testing against adaptive attackers. Participants will leave with three practical questions to eval ...