AI Agent Threat Matrix

August 31, 2026 · View on GitHub

A structured framework for classifying, detecting, and defending against attacks on AI agent systems.

Version 1.1 | June 2026 | Threats Matrix | OpenA2A


Purpose

AI agents operate differently from traditional software. They make decisions based on natural language, delegate actions to tools with varying trust levels, communicate with other agents via open protocols, and maintain persistent memory that can be poisoned. These properties create attack surfaces that existing frameworks do not adequately model.

The AI Agent Threat Matrix classifies attacks against AI agent systems into 9 tactics and 61 techniques, organized by kill chain stage. Every technique carries an explicit evidence tier: observed in the wild, validated in a controlled lab environment, or adapted from a traditional environment and marked as such. Every technique maps to automated detection and a defensive control; 60 of the 61 also carry a reproducible lab scenario.

What This Covers (and What It Doesn't)

This matrix covers the agent layer — the infrastructure between the model and the user:

  • Governance file manipulation (SOUL.md, system prompts, behavioral constraints)
  • Skill and plugin supply chain attacks
  • MCP and A2A protocol exploitation
  • Agent memory poisoning and persistence
  • Credential exposure through agent infrastructure
  • Sandbox escape via framework defaults
  • Cross-agent lateral movement and identity attacks

This matrix does not cover:

  • Model-level attacks (adversarial examples, training data poisoning) — see MITRE ATLAS
  • Prompt injection as a standalone topic — see OWASP Top 10 for LLM
  • Traditional enterprise network attacks — see MITRE ATT&CK

The Agent Threat Matrix is designed to work alongside these frameworks, not replace them.


Matrix Overview

TacticKill Chain StageTechniquesDescription
Reconnaissance17Map the target agent's attack surface, capabilities, and behavioral boundaries
Initial Access29Gain control over agent behavior through prompt manipulation or input exploitation
Credential Harvest36Extract API keys, tokens, and credentials from agent context and connected services
Privilege Escalation47Escalate capabilities beyond declared scope or bypass authorization
Lateral Movement56Pivot from compromised agent to connected services or other agents
Persistence67Establish persistent access surviving restarts and session changes
Collection77Gather and stage data from databases, file systems, and APIs
Exfiltration86Transfer collected data out of target environment
Impact96Modify data, deploy malicious code, or disrupt services

61 techniques across 9 tactics. 40 attack classes grouping related techniques. 16 techniques with real-world evidence. 42 techniques validated in controlled lab environments. 3 techniques adapted from traditional environments (marked as such).


Evidence Standard

Every technique in this matrix is assigned an evidence tier:

TierMeaningCount
ObservedConfirmed in real-world production systems16 (26%)
ValidatedReproduced in controlled lab environment (DVAA)42 (69%)
AdaptedWell-understood traditional technique applied to agent context, not yet observed agent-specifically3 (5%)

We do not publish purely theoretical techniques. Every entry has either a real-world observation, a reproducible lab scenario, or an established traditional precedent.

See EVIDENCE_AUDIT.md for the full justification of every technique's evidence tier.


Cross-Framework Mapping

This matrix maps every technique to four external references:

FrameworkWhat It Provides
HackMyAgentAutomated detection (310 static checks, 164 attack payloads)
DVAALab validation (21 vulnerable agents, 12 vulnerability categories)
OASBDefensive controls (46 controls across 10 domains)
MITRE ATT&CK / ATLAS / OWASP LLMAn index of which techniques this repository's mapping documents name against each framework, and which they do not

See cross-references/ for detailed mapping documents.


Attack Classes (40)

Attack classes group related techniques by the underlying vulnerability pattern:

Governance (11 classes)

ClassDescriptionTechniques
SOUL-POISONMalicious instructions injected into governance files at write-timeT-2001, T-2003, T-2007, T-2008
SOUL-DRIFTMulti-turn sequences gradually eroding behavioral boundariesT-2004, T-4006
SOUL-INJECTConflicting instructions via tool outputs or indirect channelsT-1003, T-2001, T-2003, T-2007, T-2008
PHANTOM-SOULAgent deployed with zero behavioral constraintsT-1006
SOUL-FORKDifferent behavior under evaluation vs productionT-5002
SOUL-HIJACKExternal content achieving constitution overrideT-4002, T-5002
SOUL-BOUNDARYExploiting ambiguous constraint definitionsT-2008
SOUL-DELEGATEDelegation without authorization chain verificationT-4001
SOUL-IMPERSONATEFalse capability claims beyond authorizationT-4002
SOUL-HVHarm avoidance override variants (4 sub-types)T-2001, T-2003
ASSEMBLY-INJECTTargeting the system-prompt assembly pipeline where components combine into exploitable injectionsT-7007

Supply Chain (11 classes)

ClassDescriptionReal-World Evidence
UNICODE-STEGOInvisible Unicode encoding instructions in source codeos-info-checker npm attack (May 2025)
MEM-POISONPersistent instructions in agent memory surviving restartsDVAA L2-04, L3-03
SKILL-MEM-AMPSkill plants payload in memory, survives uninstallDVAA validated
RAG-POISONMalicious content in vector databases retrieved by RAG pipelineAcademic research, DVAA RAGBot
HEARTBEAT-RCEPeriodic instruction fetch via heartbeat URL persistenceOpenClaw heartbeat mechanism, NemoClaw H-007
SKILL-FRONTMATTERYAML metadata injection bypassing content filtersDVAA PluginBot
SKILL-EXFILSkill exfiltrates data outside declared tool boundariesDVAA tool chain scenarios
ORG-SKILL-SPREADCompromised admin skill propagates organization-wideClawHavoc campaign patterns
SUPPLY-CHAIN-INSTALLUnsigned installation scripts executed without integrity verification — curl|sh without checksum
FAKETOOL-INJECTMCP tool impersonation, squatting, and schema poisoning attacks
PERSIST-STATECross-session persistence via memory poisoning, state tampering, and cached context injection

Infrastructure (10 classes)

ClassDescriptionReal-World Evidence
GATEWAY-EXPLOITMisconfigured API gateways exposing agent infrastructure~75K OpenClaw gateways unauthenticated
MCP-EXPLOITMCP server config, tool permissions, transport securityDVAA MCP agents, MCP protocol analysis
RETROACTIVE-PRIVExisting credentials silently gaining AI permissions32 API keys in HTTP responses (Jan 2026)
LLM-EXPOSELLM inference endpoints exposed without authentication~56K Ollama instances (Mar 2026)
AITOOL-EXPOSEAI development tools exposed (Jupyter, MLflow, Gradio)~8.3K Jupyter, ~740 MLflow (Mar 2026)
CODE-INJECTIONCommand injection via unsanitized inputsNemoClaw C-001, C-002
INTEGRITY-BYPASSDigest/hash bypass on empty or missing valuesNemoClaw C-005
TOCTOU-RACETime-of-check-time-of-use race in verification pipelinesNemoClaw C-006
A2A-EXPOSEAgent-to-Agent protocol endpoints publicly discoverable without access control
PARSER-DIFFERENTIALExploits differences between parser implementations to bypass security controls

NemoClaw-Specific (5 classes)

ClassDescriptionEvidence
NEMO-CRED-LEAKCredential exposure in NemoClaw configurationC-004 (CLI args), H-004 (env passthrough)
NEMO-NETWORK-EXPOSENetwork services bound to public interfacesC-004 (gateway), C-005 (k3s)
NEMO-SUPPLY-CHAINSupply chain integrity bypassC-003 (curl|sh), C-005 (digest bypass)
NEMO-SANDBOX-ESCAPESandbox isolation failureH-001 (Docker privileged), H-004 (Landlock)
NEMO-OPENCLAW-INHERITInherited OpenClaw flaws surviving sandboxingH-007 (Telegram pre-allowed)

Identity (2 classes)

ClassDescription
AGENT-IMPERSONATEFalse capability claims in A2A communications
BEHAVIORAL-IMPERSONATEStolen credentials detected via behavioral baseline mismatch

Sandbox (1 class)

ClassDescription
SANDBOX-ESCAPEGeneral sandbox escape via privileged containers or LSM degradation

How to Use This Framework

For Red Teams

Follow the kill chain stages sequentially. Use technique IDs to plan attack paths. Reference DVAA challenges for practice. The Attack Paths section provides complete worked examples.

For Blue Teams

Map your defenses against each tactic. Use OASB controls as a checklist. Any tactic without detection or prevention represents a gap. Focus on breaking the chain at the earliest stage.

For Researchers

Cite techniques using their IDs (e.g., "ATM T-2001"). The catalog is extensible — contribute new techniques via pull request with evidence requirements.

For Vendors

Map your product's detection capabilities to ATM technique IDs. This enables customers to understand which agent-specific threats your product covers.


Attack Paths

Complete kill chain traversals demonstrated in DVAA:

Path A — API Agent Full Compromise: T-1001 → T-2001 → T-3001 → T-5004 → T-7001 → T-8002 (LegacyBot → ToolBot: recon, inject, harvest creds, pivot, enumerate files, exfiltrate)

Path B — Memory Persistence Chain: T-1001 → T-2001 → T-6001 → T-7004 → T-8005 (MemoryBot: recon, inject, persist in memory, dump memory, exfiltrate via conversation)

Path C — Multi-Agent A2A Chain: T-1006 → T-2001 → T-4002 → T-5002 → T-9001 (Orchestrator → Worker → ToolBot: discover agent card, inject, impersonate admin, pivot via A2A, modify data)

Path D — Supply Chain to Full Compromise: T-1002 → T-2005 → T-6004 → T-5003 → T-9006 (PluginBot → ProxyBot: discover tools, inject via tool description, backdoor skill, hop MCP servers, compromise downstream)


Repository Structure

agent-threat-matrix/
├── README.md                 # This document
├── EVIDENCE_AUDIT.md         # Evidence tier justification for every technique
├── matrix.json               # Machine-readable matrix (full data)
├── tactics/                  # One file per tactic (kill chain stage)
├── techniques/               # One file per technique (T-XXXX)
├── attack-classes/           # One file per attack class
├── cross-references/         # MITRE ATT&CK, ATLAS, OWASP LLM mappings
├── CONTRIBUTING.md           # How to propose new techniques
├── CHANGELOG.md
└── LICENSE                   # Apache-2.0

canonical-classes.json maps the canonical attack classes to the technique ids in matrix.json, and is validated by scripts/check_canonical_classes.py.


How to Cite

When referencing individual techniques:

AI Agent Threat Matrix T-2001 (Direct Prompt Injection). OpenA2A, 2026. https://threats.opena2a.org/techniques/T-2001

When referencing the framework:

OpenA2A. "AI Agent Threat Matrix v1.1." June 2026. https://threats.opena2a.org


Contributing

See CONTRIBUTING.md for the full process. In summary:

  1. Every new technique must have either a real-world observation or a reproducible lab scenario
  2. Assign the next available technique ID in the appropriate stage range
  3. Include HMA detection check ID, DVAA validation (if applicable), and OASB control mapping
  4. Submit as a pull request with evidence documentation

License

Apache-2.0. The framework is free to use, cite, and build upon.


ProjectRole
HackMyAgentAutomated detection for ATM techniques (310 static checks, 164 payloads)
DVAALab validation environment (21 agents, 12 vulnerability categories)
OASBDefensive benchmark (46 controls)
AI Agent Kill ChainTactical framework (the 9-stage progression model)

Maintained by OpenA2A. Contributions welcome via pull request.