Evidence library / Communities and Working groups

MITRE ATLAS

Adversarial threats to AI systems: tactics, techniques, and case studies, including agentic-system techniques. ATLAS catalogs adversarial techniques, mitigations and case studies. It is not a conformance checklist.

Last reviewed:

This page describes AAES Govern’s controls. AAES Operate has private-work permissions, operational reviews and recorded-spend limits; it does not include Govern. About AAES Govern

MITRE ATLAS catalogs ways attackers can misuse AI systems. This page explains selected threats and the AAES Govern controls designed against them. It is a discussion of design themes, not a measured coverage assessment.

These pages describe the product. They are not a certification or a legal opinion.

Reading this note

ATLAS is authoritative for its own tactics, techniques, and case studies. This note groups selected design themes in AAES's own words; it is not an edition-pinned, technique-by-technique mapping. A reproducible coverage assessment would need exact technique references, an identified catalog revision and tests against the evaluated deployment. AAES claims no coverage percentages it has not measured.

Instrument identity

Title
MITRE ATLAS (Adversarial Threat Landscape for Artificial-Intelligence Systems)
Issuing body
MITRE
Instrument type
Adversarial threat knowledge base: tactics, techniques, and case studies, including agentic-system techniques such as prompt injection, tool misuse, and agent orchestration abuse. A threat catalog, not a control framework.
Official source
atlas.mitre.org

Question

For the agentic technique families ATLAS names, which AAES Govern controls are designed against them?

For reviewers using ATLAS: pre-execution interception, intent-aligned policy, and authorization decisions (rows R1, R3 and R4 of the AARM/CS 1 draft catalog AAES maps, whose numbering differs from public AARM v1.0) are the controls designed against the tool-misuse and action-execution techniques. A caller-declared provenance status (asserted or unknown, R11) can raise the required approval tier through policy; AAES does not inspect content or verify the source. These controls constrain tool actions. They do not detect attacks in model input.

Selected contribution

Selected design themes informed by ATLAS, with AAES Govern control boundaries
Selected design themeWhat a reviewer can examine in AAES GovernBoundary
Tool misuse and action executionAgents can discover callable tools only within their current permissions; the capability catalog may show other tools they can request permission to use; actions classified as irreversible require approval by an authorized person; an agent cannot approve itself; recorded decisions are sealed in a hash-chained record.Enforcement depends on control of the agent's credential path. An injected action that satisfies the configured policy may still be authorized.
Untrusted content and prompt injectionA caller-declared provenance status (asserted or unknown) can raise the required approval tier through policy. AAES does not inspect content or verify the source. An injected instruction that tries to become an action still has to pass the same decision as a legitimate one.AAES Govern does not detect or prevent prompt injection in model input. It constrains what a routed action can do, whether the instruction was legitimate or injected.
Agent orchestration abuseDelegation chains with verified per-hop narrowing.AAES Govern has no behavioral analytics. Review recorded decisions with your own detection tools. Activity outside AAES is visible only where telemetry is supplied.

Preconditions and gaps

  • A catalog, not a checklist. Mapping to ATLAS does not establish coverage, and AAES publishes no coverage percentages it has not measured.
  • Techniques drift. These themes were last checked against ATLAS on 4 November 2026, when the latest data release was v2026.10 (28 October 2026). The note names technique families, not technique IDs, so it is not pinned to that release. ATLAS publishes data releases about monthly, and a later release can rename or move techniques.
  • Credential-path control is required. Activity outside AAES Govern is visible only where telemetry is supplied. Observation is not enforcement.
  • AAES is not a detection product. No SOC 2 report or penetration-test report exists. AAES has a documented scope and test plan for an independent penetration test. See security disclosures.

For implementation and testing details beyond this note, use Security and implementation posture and the evaluation page. Results from a client's test tenant are produced during a scoped evaluation.

Client responsibility

The organization runs its own threat modeling against ATLAS. The client remains the regulated entity.

Choose the tool actions to test, identify who holds their credentials, and list the questions the decision records should answer.

Free workflow assessment