Running an Agentic Threat Model
An editorially reviewed technical article by Ravindra Annam on threat modeling agents as they act across identities, tools, memory, and delegated authority.
Read the published article.
ACM Queue · September 2026 issue
DOI: 10.1145/3848521
This article examines agentic workflows through the relationship between instructions, authority, and execution. The practical question is what an agent can cause to happen after it has read context, selected a tool, or delegated a task.
Questions to bring to a design review
- Which identity performs a consequential action?
- What permission reaches the tool or downstream service?
- Can untrusted content alter the target, scope, or intended result?
- What evidence shows that the receiving system enforced the limit?
The article is technical writing reviewed through ACM Queue's editorial process. It is listed separately from scholarly journal papers and completed peer-review service.
Review a workflow with your team.
How to threat model an AI agent
Map inputs, identities, delegated authority, and tool actions. Turn abuse paths into tests.
Read the guide →Agent Threat Model Worksheet
Record actors, boundaries, abuse paths, test evidence, and a review decision.
Download worksheet →Find the connected publication record
Explore the DOI and author-linked profiles without counting multiple records as additional publications.
View research records →