Most teams use MITRE ATLAS to describe how AI systems are attacked, then stop before the part that tells them what to do about it. ATLAS ships a mitigations catalogue linked to its techniques, and as of release 2026.09 (15 September 2026) it holds 40 mitigations, from AML.M0000 Limit Public Release of Information to AML.M0039 AI Honeypots, covering predictive models, generative AI and autonomous agents. This guide is the control side of ATLAS: every mitigation in one reference table, how to choose among them for your type of AI system, where the catalogue has gaps that ATT&CK and general security controls must fill, and how to roll the controls out and measure them. It is for AI security leads, threat modellers, red teams and architects who already know the ATLAS tactics. The framework itself is introduced in the MITRE ATLAS Guide.
Key takeaways
- The catalogue changes monthly. ATLAS publishes content releases named by month; release 2026.07 renamed many mitigations and added AML.M0035 AI Red Team, 2026.08 added agent authority and scope controls, and 2026.09 added AI Honeypots. Pin the release you map against.
- Three kinds of control. Each mitigation is tagged Policy, Technical - AI or Technical - Cyber; in 2026.09 there are 21 Technical - AI, 14 Technical - Cyber and 8 Policy tags (a few mitigations carry more than one).
- Agent controls are now a large block. AML.M0026 to AML.M0033 and AML.M0036 to AML.M0038 deal with agent permissions, tools, memory, segmentation, resource limits and scope, which is where most new techniques are landing.
- Coverage is incomplete by design. In the 2026.09 data, 74 of 208 technique and sub-technique entries have no mitigation linked, including Unsecured Credentials (AML.T0055) and Credentials from AI Agent Configuration (AML.T0083); fill those gaps from ATT&CK and your existing control set.
- A mitigation is a category, not a product. "Generative AI Guardrails" or "AI Telemetry Logging" describe outcomes; you still need to specify configuration, owners, tests and metrics before they reduce risk.
How the mitigations catalogue is structured
ATLAS data is published in the mitre-atlas/atlas-data GitHub repository. Since release 2026.05 content versions (for example 2026.09) are separate from the data format version (6.0.0), and dist/ATLAS-latest.yaml always points to the newest content. The older dist/ATLAS.yaml file is deprecated and stopped at release 2026.04, so tools and spreadsheets built on it are missing every mitigation from AML.M0035 onwards and show old names for renamed ones.
- Mitigation object: an ID (AML.M0000 to AML.M0039), name, description, categories and the AI lifecycle phases where it applies.
- Relationships: each "mitigates" relationship links a mitigation to a technique with a sentence explaining how it helps, which is the most useful text when writing control requirements.
- Lifecycle phases: Business and Data Understanding, Data Preparation, AI Model Engineering, AI Model Evaluation, Deployment, and Monitoring and Maintenance. Deployment is the most common tag (28 of 40), reflecting the shift toward runtime and agent controls.
- Renames to watch: AML.M0002 is now Predictive AI Output Obfuscation (formerly Passive AI Output Obfuscation), AML.M0004 is now Limit AI Service Query Volume and Rate (formerly Restrict Number of AI Model Queries), and AML.M0003, AML.M0006, AML.M0009, AML.M0010 and AML.M0015 now carry a "Predictive AI" prefix that makes clear they target classical models rather than LLMs.
The full ATLAS mitigations reference (release 2026.09)
| ID | Mitigation | Category | Main techniques limited |
|---|---|---|---|
| AML.M0000 | Limit Public Release of Information | Policy | Reconnaissance of the AI stack, system prompt and agent configuration discovery |
| AML.M0001 | Limit Model Artifact Release | Policy | Acquiring public models and data, proxy model creation |
| AML.M0002 | Predictive AI Output Obfuscation | Technical - AI | Model extraction, inversion, membership inference (AML.T0024) |
| AML.M0003 | Predictive AI Model Hardening | Technical - AI | Adversarial examples, evasion (AML.T0015, AML.T0043) |
| AML.M0004 | Limit AI Service Query Volume and Rate | Technical - Cyber | Extraction, cost harvesting, denial of AI service |
| AML.M0005 | Control Access to AI Models and Data at Rest | Policy | Poisoning, model tampering, artifact collection, IP theft |
| AML.M0006 | Predictive AI Ensembles | Technical - AI | Evasion and transfer attacks |
| AML.M0007 | Sanitize Training Data | Technical - AI | Training data poisoning (AML.T0020) |
| AML.M0008 | Validate AI Model | Technical - AI | Backdoors, poisoned models, data leakage |
| AML.M0009 | Predictive AI Multi-Sensor Fusion | Technical - Cyber | Physical attacks on sensors, deepfakes |
| AML.M0010 | Predictive AI Input Restoration | Technical - AI | Adversarial perturbations |
| AML.M0011 | Restrict Library Loading | Technical - Cyber | Unsafe model files such as pickle, malicious packages |
| AML.M0012 | Encrypt Sensitive Information | Technical - Cyber | Artifact collection, IP theft |
| AML.M0013 | Code Signing | Technical - Cyber | Supply chain compromise of AI software and models |
| AML.M0014 | Verify AI Artifacts | Technical - Cyber | Tampered models and datasets |
| AML.M0015 | Predictive AI Adversarial Input Detection | Technical - AI | Adversarial inputs, resource exhaustion |
| AML.M0016 | Vulnerability Scanning | Technical - Cyber | Unsafe artifacts, exploitation of AI software |
| AML.M0017 | AI Model Distribution Methods | Policy | Full model access on edge devices, IP theft |
| AML.M0018 | User Training | Policy | Malicious artifacts, phishing, deepfakes |
| AML.M0019 | Control Access to AI Models and Data in Production | Policy | Inference API abuse, valid accounts, account creation, token reuse |
| AML.M0020 | Generative AI Guardrails | Technical - AI | Prompt injection, jailbreak, data leakage, RAG poisoning, malicious commands |
| AML.M0021 | Generative AI Guidelines | Technical - AI | Prompt injection, jailbreak, system prompt extraction |
| AML.M0022 | Generative AI Model Alignment | Technical - AI | Jailbreak, harmful outputs, malicious command generation |
| AML.M0023 | AI Bill of Materials | Policy | AI supply chain compromise |
| AML.M0024 | AI Telemetry Logging | Technical - Cyber | Exfiltration via inference API, prompt injection, agent tool abuse |
| AML.M0025 | Maintain AI Dataset Provenance | Technical - AI | Poisoning, dataset integrity erosion |
| AML.M0026 | Privileged AI Agent Permissions Configuration | Technical - Cyber | Agent tool invocation, data and credential collection from RAG, data destruction |
| AML.M0027 | Single-User AI Agent Permissions Configuration | Technical - Cyber | Same techniques, for agents acting on behalf of one user |
| AML.M0028 | AI Agent Tools Permissions Configuration | Technical - Cyber | Abuse of tools shared across agents |
| AML.M0029 | Human In-the-Loop for AI Agent Actions | Technical - AI | Exfiltration and destruction via agent tools |
| AML.M0030 | Restrict AI Agent Tool Invocation on Untrusted Data | Technical - AI | Indirect prompt injection triggering tools |
| AML.M0031 | Memory Hardening | Technical - AI | Agent context and memory poisoning (AML.T0080) |
| AML.M0032 | Segmentation of AI Agent Components | Technical - Cyber | Tool abuse, credential harvesting from tools, active scanning |
| AML.M0033 | Input and Output Validation for AI Agent Components | Technical - AI | Direct, indirect and triggered prompt injection |
| AML.M0034 | Deepfake Detection | Technical - AI | Generated media in phishing and identity verification |
| AML.M0035 | AI Red Team | Policy, Technical - AI, Technical - Cyber | Linked to 33 techniques, the widest coverage in the catalogue |
| AML.M0036 | Limit AI Workload Resource Consumption | Technical - AI | Denial of AI service, cost harvesting, agentic resource consumption |
| AML.M0037 | AI Agent Authority Expansion Controls | Technical - AI | Autonomous reconnaissance, attack-path adaptation and orchestration |
| AML.M0038 | AI Agent Scope Drift Detection | Technical - AI | Same autonomous techniques (AML.T0116, AML.T0117, AML.T0124) |
| AML.M0039 | AI Honeypots | Technical - AI, Technical - Cyber | Reconnaissance, scanning, exposed AI services, agent capability discovery |
Choosing mitigations by AI system type
Not every mitigation applies to every system. The "Predictive AI" mitigations target classifiers and detectors; the agent mitigations only matter once a model can call tools. Start from the system type, then refine with your threat model.
| System type | Start with | Add when exposed externally | Usually not relevant |
|---|---|---|---|
| Predictive model behind an API (fraud score, malware classifier) | AML.M0005, AML.M0019, AML.M0024, AML.M0008 | AML.M0002, AML.M0004, AML.M0015, AML.M0003 | Agent permission controls |
| Generative AI chat or RAG application | AML.M0019, AML.M0020, AML.M0021, AML.M0024, AML.M0036 | AML.M0004, AML.M0000, AML.M0039 | Predictive model hardening and ensembles |
| Single-user assistant agent (acts for one person) | AML.M0027, AML.M0030, AML.M0033, AML.M0029, AML.M0031 | AML.M0032, AML.M0036 | Multi-sensor fusion |
| Privileged or multi-user agent (service desk, operations, coding agent with deploy rights) | AML.M0026, AML.M0028, AML.M0032, AML.M0029, AML.M0024 | AML.M0037, AML.M0038, AML.M0035 | Predictive input restoration |
| Model development and supply chain | AML.M0014, AML.M0013, AML.M0023, AML.M0011, AML.M0016 | AML.M0025, AML.M0007 | Runtime guardrails |
| Edge or on-device model | AML.M0017, AML.M0012, AML.M0001 | AML.M0002 | Agent controls unless the device runs an agent |
Small versus large organisations: a small team consuming a hosted model should focus on the deployment-phase controls it can actually own (access, logging, guardrails, agent permissions, resource limits) and rely on supplier assurance for model engineering controls. A model developer or large enterprise with its own training pipeline needs the data preparation and evaluation controls as well, plus a standing AI red team (AML.M0035).
Where identity fits
Several of the highest-coverage mitigations are identity controls under another name. AML.M0019 requires users to verify identity before accessing a production model and covers valid accounts, account creation and use of application access tokens. AML.M0026 and AML.M0027 call for RBAC, ABAC and least privilege for agents, and AML.M0027 explicitly covers lifecycle management: establishing the agent's identity and managing its access over time. AML.M0037 requires that the maximum authority available to an agent is granted before runtime, enforced outside the agent, and not dependent on system prompts or model alignment. The PocketOS database deletion shows what is at stake: a coding agent found an account-scoped token in an unrelated file and deleted a production database and its backups in one API call, with no approval step. In practice these mitigations are delivered through your NHI programme: see the Agentic AI Identity Guide and the AI Agent Authorisation Guide.
Filling the coverage gaps
In the 2026.09 data, 74 technique entries have no ATLAS mitigation linked. Many are reconnaissance and resource development steps that defenders cannot prevent directly, but some are squarely in scope. Borrow controls from ATT&CK and your baseline.
| Uncovered ATLAS technique | Practical control | Borrowed reference |
|---|---|---|
| AML.T0055 Unsecured Credentials | Remove secrets from code, notebooks and agent workspaces; vault and rotate | ATT&CK M1026 Privileged Account Management, M1027 Password Policies |
| AML.T0083 Credentials from AI Agent Configuration | Keep API keys out of MCP and agent config files; use short-lived tokens injected at runtime | ATT&CK M1026; NIST SP 800-53 IA-5 |
| AML.T0110 AI Agent Tool Poisoning | Allow-list tools and MCP servers, pin versions, review tool descriptions | ATT&CK M1048 Application Isolation and Sandboxing; ATLAS AML.M0014 by analogy |
| AML.T0103 Deploy AI Agent | Inventory and approve agents; block unapproved agent runtimes on endpoints | NIST SP 800-53 CM-7 Least Functionality |
| AML.T0065 LLM Prompt Crafting | Not preventable at source; rely on downstream guardrails and logging | ATLAS AML.M0020, AML.M0024 |
The MITRE ATT&CK Guide for Defenders covers the ATT&CK mitigation set used above.
Implementing ATLAS mitigations
- Pin a release and build the map. Load the current ATLAS data (not the deprecated file), record the release, and list the techniques your threat model says are relevant to each AI system.
- Select mitigations per system using the table above, and record which are delivered by you, by a supplier and by the platform.
- Write each as a testable requirement. For example, AML.M0036 becomes "agent workflows are limited to N tool calls, M retries and a spend ceiling per task, enforced by the orchestrator"; AML.M0030 becomes "tools with write or external side effects require confirmation once retrieved web or email content is in context".
- Deploy in monitor mode where possible. Run guardrails and scope drift detection in log-only mode for a pilot group, review false positives, then enforce. Keep the previous configuration so you can roll back quickly.
- Validate with red teaming. Use AML.M0035 exercises mapped to the techniques each mitigation claims to limit, and record which attacks still succeed.
- Review on each release. Check new and renamed mitigations monthly or quarterly and update the map.
Measuring coverage and effectiveness
- Mapping coverage: percentage of in-scope ATLAS techniques with at least one implemented mitigation per AI system.
- Validated coverage: percentage of implemented mitigations tested by red team or automated evaluation in the last quarter.
- Agent least privilege: percentage of agents with dedicated, scoped identities and documented maximum authority (AML.M0026, AML.M0027, AML.M0037).
- Telemetry completeness: percentage of AI systems logging prompts, outputs, tool calls and agent identity (AML.M0024).
- Release currency: age of the ATLAS release your map is based on (target under three months).
Common failure patterns
- Mapping against stale data: spreadsheets built from the deprecated ATLAS file miss the agent, red team and honeypot mitigations and use pre-2026.07 names.
- Treating guardrails as the whole answer: AML.M0020 is linked to many techniques, but it evaluates content; it does not replace agent permissions, segmentation or approval for consequential actions.
- Enforcing limits inside the prompt: AML.M0037 states that authority expansion controls must sit outside the agent and not rely solely on system prompts or model alignment.
- Claiming a mitigation without a test: "we have logging" without evidence that tool calls and agent identity are captured.
- Ignoring the gaps: assuming ATLAS coverage is complete, so unsecured credentials in agent configuration go unaddressed.
Lessons from incidents
ATLAS case studies show which mitigations would have changed outcomes. In the GTG-1002 Claude Code espionage campaign (AML.CS0069), attackers hid intent behind a false security-testing persona, so model-side controls (AML.M0022) were bypassed and provider telemetry and account controls (AML.M0024, AML.M0019) did the detection work. In the model distillation campaigns against Claude (AML.CS0056), per-account query limits (AML.M0004) were evaded through thousands of accounts, which is why the 2026.07 description of AML.M0004 now tells defenders to monitor for distributed requests and account rotation. The PocketOS deletion linked above is a textbook case for AML.M0026 and AML.M0029: scoped credentials and an approval gate for destructive actions.
Framework mapping
MITRE does not publish an official crosswalk from ATLAS mitigations to NIST controls. The mapping below is NHI Mgmt Group's own and is indicative.
| ATLAS mitigations | NIST SP 800-53 Rev. 5 | NIST CSF 2.0 | ATT&CK mitigation |
|---|---|---|---|
| AML.M0005, AML.M0019, AML.M0026, AML.M0027, AML.M0028 | AC-3, AC-6, IA-2 | PR.AA-05 | M1026, M1032 |
| AML.M0024 | AU-2, AU-12, SI-4 | DE.CM-09 | M1047 Audit |
| AML.M0032 | SC-7, SC-39 | PR.IR-01 | M1030, M1048 |
| AML.M0013, AML.M0014, AML.M0023 | SR-4, SI-7 | GV.SC-05 | M1051 Update Software (partial) |
| AML.M0035 | CA-8 Penetration Testing | ID.RA-01 | None direct |
| AML.M0029 | AC-5 Separation of Duties | PR.AA-05 | None direct |
Practitioner checklist
- Load ATLAS from
dist/ATLAS-latest.yamland record the release number in your mapping. - Classify each AI system as predictive, generative, single-user agent, privileged agent, pipeline or edge.
- Select mitigations per system and assign an owner and delivery party for each.
- Write each mitigation as a testable requirement with configuration values.
- Implement agent identity, permission and authority controls outside the model.
- Turn on AI telemetry logging that captures prompts, outputs, tool calls and agent identity.
- Cover the gaps (unsecured credentials, agent configuration secrets, tool poisoning) with ATT&CK and baseline controls.
- Validate mitigations with AI red team exercises mapped to techniques.
- Review the map against each new ATLAS release at least quarterly.
Standards and references
- MITRE ATLAS: Mitigations
- MITRE ATLAS: atlas-data repository
- MITRE ATLAS: AI Red Team
- MITRE ATLAS: AI Agent Authority Expansion Controls
- MITRE ATT&CK: Privileged Account Management
- NIST: SP 800-53 Rev. 5
- NIST: Cybersecurity Framework 2.0
Related NHI Mgmt Group resources: MITRE ATLAS Guide · AI Agent Threat Modelling Guide · Model Extraction and Distillation Defence Guide · OWASP Top 10 for Agentic Applications Guide