Skip to main content

Framework-coverage matrix

Honest framing. This page replaces the older README claim of “MITRE ATLAS v5.4.0 mappings” (which read as full-catalogue coverage). What AgentGuardian actually ships, today, against the corpus on disk: 11+ MITRE ATLAS techniques covered (see the matrix below for the exact set - currently 26 techniques). The remaining ~85% of the v5.4.0 catalogue is out of scope for a black-box agent scanner - training-pipeline compromise, ML-platform-internal techniques, and infrastructure-layer attacks do not surface at the agent’s I/O boundary and are therefore not probed.
All three tables below are auto-generated from 124 probe YAMLs by scripts/build_coverage_matrix.py. Re-run that script after adding or removing probes - the doc is the loader’s view of the world, not a hand-maintained narrative.

OWASP Top 10 for Agentic Applications 2026

Auto-generated from src/agent_guardian/probes/asiNN/*.yaml — one row per ASI category. All ten categories are present in the shipped corpus.
ASINameProbe countExample probe IDs
ASI01Goal Hijack20ASI01-GH-001, ASI01-GH-002, ASI01-GH-003
ASI02Tool Misuse8ASI02-TA-001, ASI02-TA-002, ASI02-TA-003
ASI03Privilege Abuse9ASI03-PII-001, ASI03-PR-001, ASI03-PR-002
ASI04Supply Chain17ASI04-SC-001, ASI04-SC-002, ASI04-SC-003
ASI05Code Execution8ASI05-CE-001, ASI05-CE-002, ASI05-CE-003
ASI06Memory Poisoning13ASI06-HITL-009, ASI06-HITL-010, ASI06-HITL-011
ASI07Agent-to-Agent (A2A)15ASI07-A2A-001, ASI07-A2A-002, ASI07-A2A-003
ASI08Cascading Failures8ASI08-CF-001, ASI08-CF-002, ASI08-CF-003
ASI09Trust Exploitation18ASI09-OH-001, ASI09-OH-002, ASI09-OH-003
ASI10Rogue Agents (drift)8ASI10-DR-001, ASI10-DR-002, ASI10-DR-003

MITRE ATLAS v5.4.0

Auto-generated from mitre_atlas: entries in every probe YAML. Rows are listed for every technique the shipped corpus actually cites (no padding, no aspirational entries). Techniques the corpus does not yet exercise are flagged honestly rather than hidden.
ATLAS TechniqueDescriptionProbe countExample probe IDs
AI Agent Context PoisoningAI Agent Context Poisoning (named, v5.4.0)18ASI01-GH-002, ASI01-GH-003, ASI01-GH-005
AML.T0012Valid Accounts8ASI03-PR-001, ASI03-PR-002, ASI03-PR-003
AML.T0024Exfiltration via ML Inference API1ASI03-PII-001
AML.T0029Denial of ML Service4ASI09-RSE-009, ASI09-RSE-010, ASI09-RSE-011
AML.T0034Cost Harvesting8ASI08-CF-001, ASI08-CF-002, ASI08-CF-003
AML.T0043Craft Adversarial Data12ASI07-A2A-013, ASI07-A2A-014, ASI07-A2A-015
AML.T0048External Harms17ASI02-TA-007, ASI04-SC-001, ASI04-SC-002
AML.T0050Command and Scripting Interpreter10ASI05-CE-001, ASI05-CE-002, ASI05-CE-003
AML.T0051LLM Prompt Injection13ASI01-GH-001, ASI01-GH-003, ASI01-GH-004
AML.T0053LLM Plugin Compromise11ASI02-TA-001, ASI02-TA-002, ASI02-TA-003
AML.T0054LLM Jailbreak15ASI01-GH-001, ASI01-GH-002, ASI01-GH-004
AML.T0058Publish Hallucinated Entities1ASI04-SC-008
AML.T0060ML Supply Chain Compromise: Hardware7ASI04-SC-001, ASI04-SC-002, ASI04-SC-003
AML.T0062Discover ML Artifacts4ASI04-SC-008, ASI06-MP-001, ASI06-MP-002
AML.T0064LLM Prompt Injection: Direct8ASI01-GH-001, ASI01-GH-002, ASI01-HCOT-001
AML.T0067LLM Prompt Injection: Indirect9ASI01-GH-002, ASI01-GH-006, ASI01-GH-008
AML.T0068LLM Prompt Self-Replication2ASI02-TA-004, ASI02-TA-008
AML.T0070RAG Poisoning2ASI01-GH-001, ASI01-GH-006
AML.T0071False RAG Entry Injection2ASI03-PII-001, ASI03-PR-007
Escape to HostEscape to Host (named, v5.4.0)6ASI05-CE-001, ASI05-CE-002, ASI05-CE-003
Exfiltration via AI Agent Tool InvocationExfiltration via AI Agent Tool Invocation (named, v5.4.0)5ASI02-TA-001, ASI02-TA-002, ASI02-TA-004
Memory ManipulationMemory Manipulation (named, v5.4.0)14ASI03-PR-007, ASI05-CE-008, ASI06-HITL-009
Modify AI Agent ConfigurationModify AI Agent Configuration (named, v5.4.0)14ASI01-GH-007, ASI03-PR-008, ASI04-SC-006
Publish Poisoned AI Agent ToolPublish Poisoned AI Agent Tool (named, v5.4.0)14ASI02-TA-007, ASI04-SC-001, ASI04-SC-002
RAG Credential HarvestingRAG Credential Harvesting (named, v5.4.0)3ASI06-MP-001, ASI06-MP-004, ASI06-MP-007
Thread InjectionThread Injection (named, v5.4.0)2ASI06-MP-002, ASI06-MP-003
Honest scope note. MITRE ATLAS v5.4.0 ships roughly 80 techniques across the full matrix. AgentGuardian is a black-box agent scanner — it can only meaningfully exercise techniques that manifest at the agent’s input / output / tool-call surface. Techniques targeting model training, data-pipeline infrastructure, or out-of-band ML-platform compromise are out of scope by design and not covered by the current corpus. Roughly 85% of the v5.4.0 catalogue falls into that out-of-scope set; the matrix above lists the in-scope subset we exercise.

CSA Agentic AI Red Teaming Guide

Auto-generated from csa_category: entries in every probe YAML. All 12 categories from the CSA Agentic AI Red Teaming Guide (Huang et al., 2025-05-28) are listed; zero-coverage rows are marked honestly rather than dropped.
CSA CategoryNameProbe countExample probe IDs
goal-instruction-manipulationGoal & Instruction Manipulation20ASI01-GH-001, ASI01-GH-002, ASI01-GH-003
authorization-control-hijackingAuthorization & Control Hijacking9ASI03-PII-001, ASI03-PR-001, ASI03-PR-002
agent-critical-system-interactionAgent ↔ Critical-System Interaction16ASI02-TA-001, ASI02-TA-002, ASI02-TA-003
supply-chain-dependencySupply Chain & Dependency17ASI04-SC-001, ASI04-SC-002, ASI04-SC-003
knowledge-base-poisoningKnowledge Base Poisoning2ASI06-MP-001, ASI06-MP-004
memory-context-manipulationMemory & Context Manipulation6ASI06-MP-002, ASI06-MP-003, ASI06-MP-005
multi-agent-exploitationMulti-Agent Exploitation15ASI07-A2A-001, ASI07-A2A-002, ASI07-A2A-003
impact-chain-blast-radiusImpact Chain & Blast Radius8ASI08-CF-001, ASI08-CF-002, ASI08-CF-003
hallucination-exploitationHallucination Exploitation14ASI09-OH-001, ASI09-OH-002, ASI09-OH-003
checker-out-of-the-loopChecker-Out-of-the-Loop5ASI06-HITL-009, ASI06-HITL-010, ASI06-HITL-011
agent-untraceabilityAgent Untraceability8ASI10-DR-001, ASI10-DR-002, ASI10-DR-003
resource-service-exhaustionResource & Service Exhaustion4ASI09-RSE-009, ASI09-RSE-010, ASI09-RSE-011

Regenerate this page with:
uv run python scripts/build_coverage_matrix.py