Skip to content

Latest commit

 

History

13 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Clinical AI Red Team

Attacks Results Tools

Adversarial testing and offensive security assessment for clinical LLM deployments. Part of the MedSecLab portfolio.

Purpose

Systematically test clinical AI security end-to-end:

Attack (CAI-00x) → Gateway → Audit Log → Detection Rule → Grafana → Wazuh

Validate that gateway controls and SIEM detections work together, then document findings in docs/red-team-report-v1.md.

Part of the MedSecLab portfolio:

Quick Start

# Terminal 1 — start gateway
cd ../clinical-ai-gateway && docker compose up

# Terminal 2 — automated campaign + validation
cd clinical-ai-redteam
./scripts/run_campaign.sh

# Terminal 3 — Garak scan
pip install -r requirements-garak.txt
./scripts/run_garak.sh

# Terminal 4 — multi-turn / PyRIT
./scripts/run_pyrit.sh

Manual demo (fewer steps):

./scripts/run-demo.sh

Roadmap

Phase Focus Status
4.1 Attack methodology (catalog, ROE, report v1) ✅ Complete
4.2 Manual payloads per attack ID ✅ Complete
4.3 Detection validation (run_campaign.sh) ✅ Complete
4.4 Garak integration ✅ Complete
4.5 PyRIT / multi-turn orchestration ✅ Complete
5 STRIDE threat model (MedSecLab) ✅ Complete

Repository Structure

clinical-ai-redteam/
├── README.md
├── campaign/
│   └── campaign-manifest.json     # Step expectations
├── docs/
│   ├── attack-catalog.md
│   ├── testing-methodology.md
│   ├── rules-of-engagement.md
│   └── red-team-report-v1.md
├── payloads/
│   ├── admin-abuse/
│   ├── baseline/
│   ├── phi-probing/
│   ├── prompt-injection/
│   └── system-prompt-extraction/
├── garak/
│   ├── configs/clinical-ai-gateway.yaml
│   ├── cai-probe-map.json
│   └── reports/                   # gitignored scan output
├── pyrit/
│   ├── scenarios/cai-005-scenarios.json
│   └── reports/                   # gitignored
├── requirements-garak.txt
├── requirements-pyrit.txt
└── scripts/
    ├── run_campaign.sh
    ├── run_garak.sh
    ├── run_pyrit.sh               # Multi-turn CAI-005
    ├── run_multiturn_campaign.py
    ├── compare_garak_campaign.py
    ├── compare_pyrit_campaign.py
    ├── run-demo.sh
    └── wazuh_offline.py

Attack Catalog

Attack ID Category Status
CAI-001 Ignore previous instructions Tested ✅
CAI-002 System prompt extraction Tested ✅
CAI-003 PHI probing Tested ✅
CAI-004 Administrative privilege abuse Tested — found & remediated (100310)
CAI-005 Multi-turn injection Tested ✅
CAI-006 Encoded injection Tested — found & remediated

Details: docs/attack-catalog.md

Documentation

Document Description
attack-catalog.md CAI IDs, payloads, Wazuh mapping
testing-methodology.md Evidence collection, pass/fail criteria
rules-of-engagement.md Lab scope and constraints
red-team-report-v1.md Findings and detection results

Garak

pip install -r requirements-garak.txt
./scripts/run_garak.sh

Probe → CAI mapping: garak/cai-probe-map.json · Details: garak/README.md

PyRIT / Multi-Turn

./scripts/run_pyrit.sh

CAI-005 scenarios: pyrit/scenarios/cai-005-scenarios.json · Details: pyrit/README.md

Works without PyRIT installed (stdlib orchestrator). Install PyRIT with pip install -r requirements-pyrit.txt when needed.

About

Adversarial testing methodology, attack simulations, and MITRE ATLAS mappings for securing clinical AI and LLM systems.

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages