AI Threat Modeling with MITRE ATLAS · Hacking Project

Systematically identify, map, and prioritize AI/ML-specific threats across the model lifecycle using MITRE ATLAS and generate actionable mitigations.

**The problem**
As enterprises deploy ML and LLM systems, traditional threat modeling (STRIDE, attack trees) fails to capture AI-specific attack surfaces such as training-data poisoning, model evasion, model extraction, and prompt injection. Security teams lack a structured, repeatable way to enumerate and prioritize threats across the ML lifecycle and map them to defensive controls. The result is unmanaged AI risk and inconsistent, ad-hoc assessments.

**What you'll build**
Build an AI threat-modeling framework and tool that takes an ML/LLM system description (data sources, training pipeline, model, deployment, and agentic tools) and systematically enumerates threats mapped to MITRE ATLAS tactics and techniques, scoring likelihood and impact and recommending mitigations. Combine a structured questionnaire/architecture-graph input with an LLM assistant that proposes ATLAS-mapped attack scenarios and generates a prioritized report with control recommendations aligned to NIST AI RMF and OWASP LLM/ML Top 10. Produce reusable threat-model templates for common AI architectures (RAG, fine-tuning, agentic).

**Objective** — Systematically identify, map, and prioritize AI/ML-specific threats across the model lifecycle using MITRE ATLAS and generate actionable mitigations.

**Expected outcome** — An AI threat-modeling tool/framework that ingests a system architecture and outputs an ATLAS-mapped threat register with risk scores, recommended controls, and a shareable assessment report.

**Difficulty** — Advanced — requires deep knowledge of the ML attack surface, threat-modeling methodology, ATLAS/NIST AI RMF, and structured risk analysis; less code-heavy but conceptually demanding.

**Primary category** — AI/ML Security (GRC/Assurance); Secondary = GRC/Risk Management, Threat Modeling, Security Architecture

**Security domains** — Primary = AI Security; Secondary = GRC, Security Research, Threat Intelligence, DevSecOps

**Technical focus** — AI threat modeling, MITRE ATLAS mapping, ML attack surface analysis (poisoning, evasion, extraction, inversion), LLM threat modeling, risk scoring, control mapping, NIST AI RMF, secure ML lifecycle, architecture analysis

**Core technologies** — Languages: Python; Frameworks/Libraries: threat-modeling libraries (pytm), graph libraries (NetworkX), FastAPI, LangChain (LLM assist); Security Tools: MITRE ATLAS knowledge base, ATLAS Navigator, OWASP ML/LLM Top 10, threat-modeling tools; Cloud/Infra: Docker, web app hosting; AI/Data: ATLAS STIX data, LLM API for scenario generation, embeddings for technique retrieval; Databases: PostgreSQL/graph DB for threat register, YAML/JSON template store

**Where it's valued** — Any AI-adopting enterprise (finance, healthcare, defense, tech), AI governance/compliance teams, security consultancies, regulated AI product vendors

**Careers it builds toward** — AI Security Architect, AI Governance/Risk Analyst, Security Architect, GRC Analyst (AI), AI Assurance Engineer, ML Security Consultant

**Framework alignment**
• MITRE ATLAS: full-framework mapping — AML.T0020 Poison Training Data, AML.T0043 Craft Adversarial Data, AML.T0015 Evade ML Model, AML.T0024 Exfiltration via ML Inference (model extraction/inversion), AML.T0051 Prompt Injection, AML.T0018 Backdoor ML Model
• NIST CSF 2.0: Govern (AI risk governance), Identify (asset and threat identification, risk assessment), Protect (control selection); aligns directly with NIST AI RMF Map/Measure/Manage
• OWASP: LLM Top 10 (full set for LLM systems) and OWASP Machine Learning Security Top 10 for classical ML threats
• CIS Controls v8: Control 3 Data Protection, Control 16 Application Software Security, Control 18 Penetration Testing (as validation of modeled threats)

**Innovation** — Improved Existing Solution / Enterprise Product — AI threat modeling is nascent with few mature tools, so a structured ATLAS+AI RMF tool has strong enterprise/GRC demand and portfolio value. Scores: Innovation 8/10, Startup 7/10, Enterprise-Readiness 8/10, Patent = No.