PhiloCyber / Open tool / Model
Threat modeling for systems that decide at runtime.
An open tool for the modeling stage of the workflow, built around how agentic systems actually fail.
Project status
In development- Status
- In development
- Version
- Not public yet
- Licence
- Not public yet
- Repository
- Not public yet
The problem
Classic threat modeling assumes deterministic components and fixed data flows. An agentic system does not work that way: the control flow is chosen by a model at inference time, the tools carry real privilege, and the trust boundaries move every time the agent delegates. Diagram-first tools capture the architecture you drew, not the paths the agent can take.
Who it is for
Security engineers reviewing an agentic architecture, AppSec teams that have to sign off on one, and builders who want to map the system before it reaches production.
What it is being built to do
This describes intended scope, not a released feature list. It can change before the first public version.
- 01/04
Describe the system as it runs: agents, the tools each one can reach, data sources, and who holds the privilege.
- 02/04
Mark the trust boundaries, including the ones only crossed at runtime through delegation.
- 03/04
Derive attack paths from the tools an agent can actually invoke, instead of from a generic checklist.
- 04/04
Produce a threat model a reviewer can disagree with, in a format that survives outside the tool.
Interface
Screenshot pending. No approved capture yet.
The tool is coming. The method already works.
Nothing here has to wait for a release. The field guide carries the same method, and the assessment shows where to apply it first.
Back to the AI Security Lab