Skip to main content
Menu

Volume I: Framework Foundations

AATMF v3 methodology, quantitative risk assessment, threat actor taxonomy, and framework architecture.

Introduction & Methodology

AI systems are probabilistic, context-dependent, and trained on human language — making them susceptible to the same manipulation techniques used against humans for millennia. Traditional cybersecurity frameworks miss the attack surfaces unique to AI: prompt injection, training data poisoning, model extraction, agentic exploitation, RAG manipulation, and the human feedback loops that shape model behavior.

AATMF fills that gap. Each technique is documented with a unique namespaced identifier, risk score (AATMF-R v3), attack procedures with example prompts, detection patterns, mitigation controls, and cross-framework references to MITRE ATLAS, OWASP, NIST, and EU AI Act.

Scope

+ Large Language Models (LLMs) and Large Reasoning Models (LRMs)
+ Multimodal models (vision, audio, video)
+ Retrieval-Augmented Generation (RAG) systems
+ Autonomous AI agents and multi-agent orchestrators
+ AI development and deployment infrastructure
+ Human-in-the-loop workflows
+ AI supply chains (models, datasets, tools, libraries)

Threat Actor Taxonomy

Actor Motivation Tactics Sophistication
Script kiddies Curiosity, clout T1, T2 Low
Bug bounty hunters Financial reward T1–T5, T10 Medium–High
Cybercriminals Financial gain T1–T3, T7–T8, T13 Medium
Corporate espionage Competitive advantage T5, T10, T13–T14 High
Nation-state actors Strategic advantage T6, T11, T13–T15 Very High
AI red teams Security improvement All Very High
Insiders Various T6, T15 Variable

Evolution

v1.0
2024

Initial framework, 8 tactics

v2.0
Late 2024

Expanded to 12 tactics, added risk scoring

v3
February 2026

20 tactics, 240 techniques, 2,152+ procedures, namespaced IDs, Volumes V–VII

AATMF-R v3 Risk Assessment

AATMF-R v3 uses a six-factor formula to quantify risk. Each factor is scored independently and combined to produce a final risk rating.

Risk = (L × I × E) / 6 × (D / 6) × R × C

Scoring Guidelines

Likelihood (L) — 1 to 5

1 Rare Requires novel research, no known PoC
2 Unlikely Requires specialized knowledge
3 Possible Known technique, moderate skill required
4 Likely Well-documented, readily available tools
5 Almost Certain Automated, commodity attack

Impact (I) — 1 to 5

1 Negligible Minor policy violation, no data exposure
2 Minor Limited harmful content, no sensitive data
3 Moderate Sensitive data exposure, service degradation
4 Major Critical data breach, safety bypass, service outage
5 Catastrophic Physical harm potential, mass data breach, systemic compromise

Exploitability (E) — 1 to 5

1 Theoretical Requires custom research and novel techniques
2 Difficult Needs deep expertise and specific conditions
3 Moderate Documented approach, some skill required
4 Easy Copy-paste attacks, minimal customization
5 Trivial Automated tools, zero skill required

Detectability (D) — 1 to 5

1 Obvious Trivially detected by basic filters
2 Easy Standard monitoring catches it
3 Moderate Requires specialized detection
4 Difficult Advanced analysis needed
5 Nearly Invisible No reliable detection method exists

Recoverability (R) — 1 to 5

1 Immediate Auto-recoverable, no intervention needed
2 Quick Simple rollback or reset
3 Moderate Requires investigation and manual remediation
4 Difficult Extended downtime, data loss possible
5 Irrecoverable Permanent damage, no full recovery path

Cost Factor (C) — 0.5 to 2.0

0.5 Minimal economic impact, internal only
1.0 Standard business impact
1.5 Significant financial or reputational damage
2.0 Catastrophic economic consequences

Interactive Calculator

Adjust the six factors below to calculate a risk score for your threat scenario. Scores vary based on deployment context — a chatbot and an autonomous financial agent score very differently on Impact and Cost Factor.

AATMF-R v3 Risk Calculator

Formula: (L × I × E) / 6 × (D / 6) × R × C

3Possible

Probability of successful exploitation

3Moderate

Severity of successful attack

3Moderate

Ease of execution

3Moderate

Difficulty of detection (5 = nearly invisible)

3Moderate

Effort to recover (5 = irrecoverable)

1Standard

Economic impact multiplier

6.8 Info
Critical
250+
High
200–249
Medium
150–199
Low
100–149
Info
0–99

Calculation

(3 × 3 × 3) / 6 × (3 / 6) × 3 × 1

= 6.8

Framework Architecture

AATMF organizes adversarial threats in a four-level hierarchy. Each level adds specificity, from strategic objectives (tactics) down to executable attack strings (prompts).

15 Tactics          → High-level adversarial objectives
└── 240 Techniques   → Specific attack methods
    ├── 2,152+ Procedures → Implementation variants
    │   └── 4,980+ Prompts  → Actual attack examples
    ├── Detection Patterns
    └── Mitigation Controls

v3 uses namespaced identifiers (T1-AT-001) that eliminate the 43 ID collisions present in earlier versions. Every identifier is globally unique with tactic membership visible at a glance.