跳转至

MLASTG → ATLAS Coverage Dashboard

Generated: v0.1 (Draft) — Data from corrected matrix with verified ATLAS IDs

Overall Coverage

                              ⬤⬤⬤⬤⬤⬤⬤⬤⬤⬤⬤⬤⬤ 72% Full
                              ⬤⬤⬤⬤ 22% Partial  
                              ⬤ 6% None
Metric Value
Total ATLAS techniques 18
Full coverage 13 (72.2%)
Partial coverage 4 (22.2%)
No coverage 1 (5.6%)
MLASVS controls mapped 64 unique controls
MLASTG test cases mapped 16 test cases (all covered)
MLASWE weaknesses connected 9 weaknesses

Coverage by Tactic

Tactic Techniques Full Partial None
Reconnaissance (TA0001) 1 0 0 1
Resource Development (TA0003) 2 1 1 0
Initial Access (TA0002) 4 3 1 0
Execution (TA0005) 2 2 0 0
Persistence (TA0006) 2 2 0 0
Credential Access (TA0007) 1 1 0 0
Discovery (TA0009) 2 2 0 0
Collection (TA0010) 2 1 1 0
ML-specific techniques 10 8 2 0
— ML Attack/Evasion 1 1 0 0
— Adversarial Perturbation 1 1 0 0
— Model Inversion 1 0 1 0
— Data Poisoning (+ sub) 3 0 3 0
— Model Extraction 1 1 0 0
— Model DoS 1 1 0 0
— LLM Prompt Injection 1 1 0 0
— LLM Data Leakage 1 1 0 0
— LLM Plugin Compromise 1 1 0 0
— ML Behavioral Manipulation 1 1 0 0

Control Density Heat Map

ATLAS ID Technique Controls Density
AML.T0010 Adversarial Perturbation 10 controls 🔥 High
AML.T0029 Model DoS 6 controls 🔥 High
AML.T0051 LLM Prompt Injection 8 controls 🔥 High
AML.T0024.002 Model Extraction 6 controls 🔥 High
AML.T0053 LLM Plugin Compromise 5 controls 🔥 Medium
AML.T0020 Data Poisoning 5 controls 🔥 Medium
AML.T0057 LLM Data Leakage 4 controls 🔥 Medium
AML.T0018 Model Inversion 3 controls ⚡ Low
AML.T0043 Input Manipulation 2 controls ⚡ Low
AML.T0056 ML Behavioral Manipulation 4 controls 🔥 Medium
— ML Model Discovery (TA0001) 0 controls ❌ None

Cross-Framework Alignment

Framework MLASTG Coverage Status
MITRE ATLAS 72% Full, 22% Partial, 6% None 🟢 Strong
OWASP LLM Top 10 10/10 fully covered 🟢 Complete
OWASP ML Top 10 10/10 fully covered 🟢 Complete
NIST AI RMF All 4 functions covered 🟢 Complete
EU AI Act GOV controls mapped 🟡 In progress

Recommendations

v0.2 (Next Release)

  1. Add Reconnaissance detection control (GOV-016a) — currently the only zero-coverage area
  2. Upgrade Data Poisoning to Full coverage with L1 automated detection tests (TEST-DATA-002)
  3. Add L1 Model Inversion prevention control (output confidence clipping)

v0.3

  1. Strengthen Backdoor prevention with automated training-time detection
  2. Add continuous base model monitoring control for supply chain
  3. Extend coverage to non-LLM output handling (AML.T0043)

v0.4+

  1. Extend to multimodal ML attacks (vision-language models)
  2. Add Reinforcement Learning controls (reward poisoning scenarios)
  3. Add Federated Learning security controls