How do I do Threat Modeling for AI agents?

Picture a triage agent that can read one repository and leave comments. On paper, it can't do much harm. But it runs as a step in a deployment pipeline, under the pipeline's cloud role, and nobody ever granted it that access.

Threat modeling an agent means finding gaps like this, between its scope and its actual reach, before an attacker does. Token Security’s Vicki Menashe starts from the identity and works outward.

You'll learn how to do threat modeling for agents by:

• Finding the identity the agent actually uses, and the roles behind it.
• Mapping the blast radius through the agent’s tools, approval settings, and triggers.
• Spotting ways the agent can raise its own privileges, including risky pairs of harmless permissions.
• Tracing attack paths from a hostile instruction to something of yours.
• Writing findings as stories that others can act on, instead of scores.

‍