Part 1 of 4 · 1 min read
Overview
Agents combine untrusted inputs, powerful tools and sensitive data in ways ordinary threat models miss. Interviewers ask "walk me through the threat model of an agent you built". A good answer draws the data-flow diagram with trust boundaries, names assets and attackers, lists top threats with controls, and says what is accepted.
About the example, and what is real
Real: the data-flow model, STRIDE-per-element enumeration, trust-gap check, AND/OR attack-tree evaluation, brute-force cheapest-cut search and the 17 test cases in code/mini-threats. Assumed: the system (a fictional support agent), node sensitivities, every leaf probability and cost, mitigation costs, and the rule that only flows toward higher trust are flagged. Not real: no real incident data, no CVSS or DREAD scoring, no MAESTRO tooling; the model flags inbound crossings only, so an outbound channel such as the email flow needs its own egress rule (module 1).