Frameworks › MITRE ATLAS

MITRE ATLAS v5.5.0

MITRE ATLAS (Adversarial Threat Landscape for AI Systems) v5.5.0 provides 167 techniques organized across 16 kill-chain tactics. ConstantX scenarios carry atlas_ids fields linking each test to the specific ATLAS techniques it exercises, giving auditors kill-chain-stage positioning for every verdict.


ATLAS Role in ConstantX Threat Modeling

ATLAS and OWASP T-codes serve different roles in ConstantX threat modeling:

The derivation chain with ATLAS:

T-code → ATLAS technique ID → Threat → ASI code → Scenario → Verdict

Coverage by Tactic

Tactic Status Key Techniques
Execution (AML.TA0005) Covered AML.T0051 (Prompt Injection), AML.T0050 (Command & Scripting), AML.T0053 (Tool Invocation), AML.T0103 (Deploy Agent)
Persistence (AML.TA0006) Covered AML.T0070 (RAG Poisoning), AML.T0080 (Context Poisoning), AML.T0080.000 (Memory)
Privilege Escalation (AML.TA0012) Covered AML.T0054 (LLM Jailbreak), AML.T0053 (Tool Invocation), AML.T0105 (Escape to Host)
Defense Evasion (AML.TA0007) Covered AML.T0068 (Prompt Obfuscation), AML.T0067 (Trusted Output Manipulation)
Credential Access (AML.TA0013) Covered AML.T0055 (Unsecured Credentials), AML.T0083 (Credentials from Config)
Exfiltration (AML.TA0010) Covered AML.T0086 (Exfiltration via Tool), AML.T0057 (LLM Data Leakage)
Impact (AML.TA0011) Covered AML.T0029 (Denial of AI Service), AML.T0101 (Data Destruction via Tool)
Initial Access (AML.TA0004) Partially covered AML.T0093 (Prompt Infiltration via Public-Facing App)
Discovery (AML.TA0008) Partially covered AML.T0069 (Discover LLM System Info), AML.T0084 (Discover Agent Config)
Collection (AML.TA0009) Partially covered AML.T0037 (Data from Local System)
Resource Development (AML.TA0003) Covered AML.T0065 (Prompt Crafting), AML.T0066 (Retrieval Content Crafting)
AI Attack Staging (AML.TA0001) Covered AML.T0043 (Craft Adversarial Data)
Command and Control (AML.TA0014) Coverable AML.T0096 (AI Service API as C2)
Lateral Movement (AML.TA0015) Not covered Requires multi-service infrastructure outside sandbox scope
Reconnaissance (AML.TA0002) Not covered Occurs before agent interaction — outside evaluation boundary

Verdict Mapping to ATLAS

ConstantX Verdict ATLAS Interpretation
valid_commit Agent completed task without triggering any ATLAS technique — no adversarial behavior observed
bounded_failure Agent attempted or was prompted toward an ATLAS technique but enforcement surfaces contained the attempt — the technique was detected and blocked
undefined_behavior Agent may have partially executed an ATLAS technique without observable enforcement — the technique may have succeeded against the evaluation target

Coverage Boundaries

Three structural limits apply to ATLAS mapping:

See ATLAS technique coverage in completed engagements
Opus 4.5 | 100% TC GPT 5.4 | 85.85% TC
Scope Your Deployment Audit

All Frameworks · OWASP ASI · NIST AI RMF · Methodology Paper