SECCourse

AI security & prompt-injection defense

Lessons10modules

Total106mfull study

Quick7mtrailer

Projects8docker labs

threat-model-workbench · STRIDE-for-LLM + ATLAS

YAML threat model + MITRE ATLAS mapper + Markdown report. Design-review checkpoint for any new LLM service.

snap/ai-security:threat-modelRepo · ai-security-threat-model

$git clonehttps://github.com/snap-dev/ai-security-threat-model.git

docker-compose.yml

services:
  workbench:
    image: python:3.11-slim
    working_dir: /app
    volumes:
      - ./src:/app/src:ro
      - ./threats:/app/threats:ro
      - ./reports:/app/reports
      - ./requirements.txt:/app/requirements.txt:ro
    environment:
      ATLAS_DB: /app/threats/atlas-techniques.json
      STRIDE_REQUIRED: "S,T,R,I,D,E"
    command: >-
      bash -c "pip install -q -r requirements.txt && python -m src.lint /app/threats/threats.yml && python -m src.report /app/threats/threats.yml --out /app/reports/threat-model.md"

Run

~/ai-security-threat-model · zsh

$ docker compose up --abort-on-container-exit

[scan] N threats parsed · [stride] coverage breakdown · reports/threat-model.md generated.

What you'll observe

threats.yml linted: every threat has stride / atlas / surface / mitigations / owner / severity

Linter fails the build if any STRIDE letter has 0 threats — forces full coverage

ATLAS references validated against vendored atlas-techniques.json

reports/threat-model.md groups threats by STRIDE / severity / owner

Output paste-ready into Confluence / Notion design review

Lift this to your work

Make this the design-review checkpoint for any new LLM service. Drop a threats.yml next to the repo; gate PRs on green threat-model run; reviewers diff threat changes alongside code. Snap requires this before any new agent / RAG ships — caught a missing tool-allow-list before the agent went live last quarter.