An agent may already have permission to open documents, send emails or change files. What happens when someone persuades it to use that access for something the organization never authorized?
Mace Ai co-founders Greg Frank and Walton Comer explore that question with TheStreet Roundtable’s James Heckman. Greg explains how manipulation across multiple interactions can push an agent beyond its assigned scope, even when it refuses a direct request, and how adversarial testing helps teams reproduce failures and gather evidence for remediation.
Walton draws a parallel with Y2K: a shared security challenge requiring coordinated work across businesses, governments and the security community.