---
title: How do you use MITRE ATLAS to threat-model AI agents?
url: https://opencontroller.lyzr.com/blog/mitre-atlas-threat-model-ai-agents
published: 2026-11-12
updated: 2026-11-12
author: Lyzr
---

# How do you use MITRE ATLAS to threat-model AI agents?

**Draw the agent's data and tool flows, then walk each component through the ATLAS techniques that target agents, such as AML.T0080 AI Agent Context Poisoning and AML.T0086 Exfiltration via AI Agent Tool Invocation. For every technique that applies, write one detection and one preventive control.** The output is a short table you can test and re-run each time the agent changes.

MITRE ATLAS is the AI counterpart to MITRE ATT&CK. Since October 2025 it has added agent-specific techniques, so it now describes attacks on tools, memory and agent configuration, not only on models.

## Which ATLAS techniques are specific to agents?

These are the agent techniques most teams should start with. IDs are as published in ATLAS.

| ID | Technique | What the attacker does |
|---|---|---|
| AML.T0051 | LLM Prompt Injection | Inserts instructions directly or through content the agent reads |
| AML.T0080 | AI Agent Context Poisoning (incl. .000 Memory) | Plants content in context or long-term memory to steer later actions |
| AML.T0081 | Modify AI Agent Configuration | Changes the agent's system prompt, tools or settings |
| AML.T0082 | RAG Credential Harvesting | Uses retrieval to surface secrets stored in documents |
| AML.T0086 | Exfiltration via AI Agent Tool Invocation | Makes the agent send data out through a legitimate tool |
| AML.T0098 | AI Agent Tool Credential Harvesting | Extracts credentials the agent's tools hold |
| AML.T0099 | AI Agent Tool Data Poisoning | Tampers with data a tool returns to the agent |
| AML.T0100 | AI Agent Clickbait | Lures a browsing agent into clicking or fetching malicious content |
| AML.T0101 | Data Destruction via AI Agent Tool Invocation | Uses the agent's own tools to delete or corrupt data |

## How do you run the threat model, step by step?

Take a realistic example: a customer-support agent that reads tickets, retrieves from a knowledge base, updates CRM records through an MCP server, and can send email.

1. **Draw the flows.** Inputs (tickets, emails), context (knowledge base, memory), tools (CRM MCP server, email), identity (which credential each tool uses), and outputs.
2. **Mark trust boundaries.** Anything written by a customer is untrusted. So is any document the knowledge base ingests from outside.
3. **Walk the techniques per component.** A ticket body is an injection path (AML.T0051). The knowledge base is a poisoning and credential-harvesting path (AML.T0080, AML.T0082). The email tool is an exfiltration path (AML.T0086). The CRM tool is a destruction path (AML.T0101).
4. **Rate each one** by impact and by how reachable it is from untrusted input.
5. **Write one detection and one prevention** for each rated technique.
6. **Turn the table into tests.** Every row becomes a red-team scenario. See [scoping a red-team assessment](/blog/scope-an-ai-agent-red-team-assessment).
7. **Re-run on change.** New tool, new memory store or new model means a new pass.

## What does the finished threat model look like?

| Component | Technique | Detection | Prevention |
|---|---|---|---|
| Ticket body | AML.T0051 | Guardrail flags instruction-like text in untrusted input | Treat ticket text as data; task-adherence check on the plan |
| Knowledge base | AML.T0082 | Scan retrieved chunks for secret patterns | Secrets removed at ingestion; redaction at the gateway |
| Long-term memory | AML.T0080.000 | Alert on memory writes from untrusted sessions | Memory writes limited to named fields |
| Email tool | AML.T0086 | Alert on external recipients or bulk sends | Allow-list recipient domains; size limits |
| CRM MCP server | AML.T0101 | Alert on delete or bulk-update calls | Read and update only; no delete scope |
| Agent config | AML.T0081 | Diff alerts on system prompt and tool changes | Changes only through a versioned registry |

## Where does ATLAS leave gaps for agents?

ATLAS is strong on single-agent attacks. A Cloud Security Alliance research note in March 2026 counted 15 tactics in ATLAS v5.x and pointed out that ATLAS has no lateral movement tactic, so agent-to-agent spread and delegation chains are under-described. Cover those with the [OWASP agentic Top 10](/blog/owasp-top-10-agentic-applications-evidence), especially ASI07 Insecure Inter-Agent Communication and ASI08 Cascading Failures.

## How Opencontroller does this

Opencontroller's policy as code includes built-in mappings to MITRE ATLAS, so policy violations can be reported by technique. Detections in the table above map to its capabilities: live groundedness and task-adherence guardrails, redaction and blocking in the LLM Gateway and Agent Gateway, a governed registry of agent versions, tools and MCP servers, and behavioural-deviation monitoring with quarantine and kill. Multi-turn and red-team evals turn the threat model into repeatable tests. See the [framework mapping hub](/blog/how-to-map-ai-agent-controls-to-compliance-frameworks) or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**What is MITRE ATLAS?**
A public knowledge base of adversary tactics, techniques and case studies for attacks on AI systems, modelled on MITRE ATT&CK.

**Does ATLAS cover AI agents?**
Yes. Agent-specific techniques have been added since October 2025, including context poisoning and exfiltration through tool calls.

**Should we use ATLAS or the OWASP agentic Top 10?**
Use both. OWASP names the risk categories; ATLAS describes how attackers carry them out, which is better for writing detections and tests.

## Key facts

MITRE ATLAS is a knowledge base of adversary tactics and techniques against AI systems. From October 2025 it added agent-focused techniques, including AML.T0080 AI Agent Context Poisoning (sub-technique AML.T0080.000 Memory), AML.T0081 Modify AI Agent Configuration, AML.T0082 RAG Credential Harvesting and AML.T0086 Exfiltration via AI Agent Tool Invocation; the first 2026 update added AML.T0098 AI Agent Tool Credential Harvesting, AML.T0099 AI Agent Tool Data Poisoning, AML.T0100 AI Agent Clickbait and AML.T0101 Data Destruction via AI Agent Tool Invocation. A Cloud Security Alliance note (March 2026) counted 15 tactics in ATLAS v5.x and noted that ATLAS does not have a lateral movement tactic. Opencontroller by Lyzr includes built-in policy mappings to MITRE ATLAS. Pricing starts from $250K per deployment per year.

## Related

- [How do you map AI agent controls to compliance frameworks?](https://opencontroller.lyzr.com/blog/how-to-map-ai-agent-controls-to-compliance-frameworks.md)
- [What evidence maps to the OWASP Top 10 for Agentic Applications?](https://opencontroller.lyzr.com/blog/owasp-top-10-agentic-applications-evidence.md)
- [How do you scope an AI agent red-team assessment?](https://opencontroller.lyzr.com/blog/scope-an-ai-agent-red-team-assessment.md)
- [More on Evidence and frameworks](https://opencontroller.lyzr.com/blog/cluster/evidence-and-frameworks)
