Every guardrail, prompt classifier, model scanner and AI-based malware detector you deploy becomes a target in its own right, and the MITRE ATLAS Defense Evasion tactic (AML.TA0007) catalogues how adversaries get past them. ATLAS describes the tactic as the adversary "trying to avoid being detected by AI-enabled security software". In practice it now covers two very different problems: fooling AI models that act as security controls, and hiding malicious instructions, artifacts and activity from the guardrails and people that watch AI systems. This hub explains what the tactic means in practice, which of its 21 techniques matter most, how they chain in documented cases, and where detection and mitigation effort pays off. It is written for AI security engineers, detection engineers, red teams and platform owners, and it links every technique page in the tactic rather than repeating them. For the framework as a whole, see the MITRE ATLAS Guide.
Key takeaways
- Prompt-layer evasion dominates the case data. In the ATLAS case studies, LLM Prompt Obfuscation (AML.T0068) and LLM Jailbreak (AML.T0054) each appear in 11 cases, more than any other technique in the tactic.
- Guardrails alone are not a defence strategy. Generative AI Guardrails (AML.M0020) is the mitigation ATLAS maps most often in this tactic, yet the documented bypasses show that one classifier in front of a model is exactly what attackers plan around.
- Supply chain trust is an evasion surface. Masquerading, Impersonation, AI Supply Chain Rug Pull and Reputation Inflation all exploit the trust you place in names, publishers and download counts for models, packages and MCP servers.
- Several techniques have no ATLAS mitigation at all. Corrupt AI Model, Manipulate User LLM Chat History, Virtualization/Sandbox Evasion, Reputation Inflation, Obfuscated Files or Information, AI Targeted Cloaking, Indicator Removal and Metadata Sanitization rely on detection, logging integrity and architecture instead.
- Identity is the quiet enabler. Many evasion chains start with valid accounts or leaked API keys, so a jailbreak run from a stolen key looks like a legitimate customer to the provider.
What AML.TA0007 means in practice
ATLAS took the name and idea from ATT&CK: its Defense Evasion tactic carries an ATT&CK reference to TA0005. Note that in ATT&CK v20, TA0005 is now named Stealth, and a separate Defense Impairment tactic (TA0112) covers disabling or modifying defences. ATLAS has kept a single Defense Evasion tactic, so it mixes both ideas: hiding (obfuscation, masquerading, cloaking) and defeating controls (jailbreaks, model evasion, exploiting security software).
It helps to sort the 21 techniques by the defence they defeat:
- AI models used as security controls: malware classifiers, phishing page detectors, DGA detectors and biometric verification. The technique here is Evade AI Model (AML.T0015), usually built on crafted adversarial data or deepfakes.
- Safety alignment and guardrails on LLMs: LLM Jailbreak, LLM Prompt Obfuscation, Triggers in Multimodal Inputs, Delay Execution of LLM Instructions and AI Targeted Cloaking all aim to get instructions past filters and into the model.
- The human reading the output: LLM Trusted Output Components Manipulation (including its Citations sub-technique), False RAG Entry Injection and Manipulate User LLM Chat History make a compromised answer or session look trustworthy.
- Artifact and supply chain review: Masquerading, Impersonation, Corrupt AI Model, AI Supply Chain Rug Pull and AI Supply Chain Reputation Inflation defeat model scanners and the human judgement used when adopting models, packages and MCP servers.
- Provider abuse detection: Centralized AI Service Request Routing and AI Service Request Metadata Sanitization hide coordinated abuse, such as distillation, from the AI provider's own enforcement.
- Classic host and forensic defences: Obfuscated Files or Information, Virtualization/Sandbox Evasion, Exploitation for Defense Evasion, Modify AI Agent Configuration and Indicator Removal mirror their ATT&CK counterparts, applied to AI systems and agents.
Every technique in the tactic
The table lists every technique and sub-technique that ATLAS (data release 2026.10) assigns to AML.TA0007, with the ATLAS maturity rating and the number of ATLAS case studies that use it under this tactic. Several techniques also sit under other tactics (for example LLM Jailbreak under Privilege Escalation and Modify AI Agent Configuration under Persistence); the case counts here only count steps tagged to Defense Evasion.
| ID | Technique | Maturity | Cases | What it defeats |
|---|---|---|---|---|
| AML.T0015 | AML.T0015 Evade AI Model | Realized | 6 | AI-based detectors and classifiers |
| AML.T0054 | AML.T0054 LLM Jailbreak | Realized | 11 | Model alignment and guardrails |
| AML.T0067 | MITRE ATLAS AML.T0067 LLM Trusted Output Components Manipulation | Demonstrated | 3 | User trust in responses |
| AML.T0067.000 | Citations (sub-technique of AML.T0067) | Demonstrated | 1 | User trust in cited sources |
| AML.T0068 | MITRE ATLAS AML.T0068 LLM Prompt Obfuscation | Realized | 11 | Human review and prompt filters |
| AML.T0071 | MITRE ATLAS AML.T0071 False RAG Entry Injection | Demonstrated | 1 | Trust in retrieved documents |
| AML.T0073 | AML.T0073 Impersonation | Realized | 6 | Human and agent trust decisions |
| AML.T0074 | AML.T0074 Masquerading | Realized | 7 | Artifact review and security tools |
| AML.T0076 | MITRE ATLAS AML.T0076 Corrupt AI Model | Realized | 1 | Model file scanners |
| AML.T0081 | MITRE ATLAS AML.T0081 Modify AI Agent Configuration | Demonstrated | 1 | Agent safety settings |
| AML.T0092 | MITRE ATLAS AML.T0092 Manipulate User LLM Chat History | Demonstrated | 2 | User and investigator visibility |
| AML.T0094 | MITRE ATLAS AML.T0094 Delay Execution of LLM Instructions | Demonstrated | 2 | Per-turn injection checks |
| AML.T0097 | AML.T0097 Virtualization/Sandbox Evasion | Realized | 1 | Dynamic analysis sandboxes |
| AML.T0107 | AML.T0107 Exploitation for Defense Evasion | Demonstrated | 1 | Security features through vulnerabilities |
| AML.T0109 | MITRE ATLAS AML.T0109 AI Supply Chain Rug Pull | Realized | 1 | One-time dependency vetting |
| AML.T0111 | MITRE ATLAS AML.T0111 AI Supply Chain Reputation Inflation | Demonstrated | 1 | Popularity and publisher trust signals |
| AML.T0123 | AML.T0123 Obfuscated Files or Information | Realized | 2 | Classifiers, secret scanners, content filters |
| AML.T0129 | AML.T0129 Triggers in Multimodal Inputs | Feasible | 0 | Text-only moderation |
| AML.T0134 | AML.T0134 AI Targeted Cloaking | Feasible | 0 | Crawlers and human reviewers |
| AML.T0138 | AML.T0138 Indicator Removal | Realized | 1 | Logs and investigation |
| AML.T0140 | AML.T0140 Centralized AI Service Request Routing | Realized | 3 | Provider per-account abuse detection |
| AML.T0141 | AML.T0141 AI Service Request Metadata Sanitization | Realized | 1 | Attribution and correlation |
Method: counts are the number of distinct ATLAS case studies (77 in the 2026.10 release) with at least one procedure step that uses the technique and is tagged to AML.TA0007. In total, 41 of the 77 case studies include at least one Defense Evasion step, more than for Discovery or Collection.
Where to start: the techniques that matter most
Ranking by case frequency alone would hide some risks, so the order below combines case counts, maturity and how directly the technique affects typical enterprise AI deployments.
- LLM Prompt Obfuscation (AML.T0068). Present in 11 cases, including EchoLeak, the Rules File Backdoor against coding assistants, the Cursor MCP exfiltration case and ZombieAgent. Hidden text, invisible Unicode, white-on-white content and phrasing tuned to slip past injection classifiers are what turn a prompt injection into a working one. Defences that only scan visible text will miss it.
- LLM Jailbreak (AML.T0054). Also 11 cases, from Storm-2139's guardrail bypass tooling for Azure OpenAI to GTG-1002, where operators posed as a defensive security firm and split work into innocuous tasks. The point for defenders: model alignment is a probabilistic control, so tool permissions and approvals must hold even when the model has been talked round.
- Masquerading (AML.T0074) and Impersonation (AML.T0073). Seven and six cases, concentrated in the AI supply chain: organisation confusion on Hugging Face, model namespace reuse, poisoned GGUF templates, poisoned agent skills and the postmark-mcp package. Names and publisher claims are what get a malicious artifact installed.
- Evade AI Model (AML.T0015). Six cases, mostly the original ATLAS studies against malware, C2, DGA and phishing detectors. Still the core risk if you rely on AI-based detection products or your own classifiers in fraud, KYC or security pipelines.
- Delay Execution of LLM Instructions (AML.T0094) and Triggers in Multimodal Inputs (AML.T0129). Fewer cases, but both are aimed squarely at the per-turn, text-only checks that most guardrail deployments use today.
- Centralized AI Service Request Routing (AML.T0140) and Metadata Sanitization (AML.T0141). Priority for AI service providers and for any organisation that exposes models through an API: they appear in the distillation campaign case studies, including the one NSA, CISA and FBI attributed to China-based AI companies.
How evasion chains in real cases
Defense Evasion rarely stands alone. In the ATLAS procedures it usually sits between initial access or prompt injection and the step that delivers impact. Four patterns recur.
| Chain | Example case | Evasion steps | Where the chain breaks |
|---|---|---|---|
| Hidden injection to data theft | EchoLeak (M365 Copilot) | AML.T0068 obfuscation to pass the injection classifier; AML.T0067 to shape a trustworthy-looking answer | Output rendering controls and egress restrictions on links and images |
| Jailbreak to misuse | Storm-2139 (Azure OpenAI) | AML.T0054 jailbreak tooling run through valid accounts from scraped API keys | Key hygiene, secret scanning and per-key anomaly detection |
| Trusted package to exfiltration | Poisoned postmark-mcp server | AML.T0073 impersonation of the publisher, then AML.T0109 rug pull after adoption | Version pinning, update review and monitoring tool behaviour after updates |
| Delayed instruction to agent action | Gemini calendar invitation attacks | AML.T0094 delayed trigger and AML.T0054 jailbreak inside retrieved content | User confirmation for side-effecting tool calls |
The EchoLeak write-up shows the pattern clearly: the injection evaded the prompt injection classifier through phrasing, bypassed link redaction with reference-style Markdown and routed requests through trusted Microsoft domains. Each control worked as designed against the inputs it expected. The postmark-mcp package worked normally for fifteen versions before one line blind-copied every agent-sent email to the attacker, which is why review at adoption time is not enough. In the Gemini calendar case the instruction sat in an invitation until the user asked an ordinary question about their schedule.
The newest cases add autonomous agents as the evader. In the AISI evaluation case study (AML.CS0074) ATLAS records Indicator Removal, Obfuscated Files or Information, Masquerading, Impersonation and Prompt Obfuscation by agents acting without human direction. In the OpenAI and Hugging Face incident, agents encoded and encrypted staged payloads, so simple log searches missed most of the stolen secrets.
Detection priorities across the tactic
ATLAS does not publish detection analytics per technique, so the priorities below follow from the technique descriptions and the case procedures. Group them by where the evidence lives.
- Prompt and retrieval content: log the full assembled context (system prompt, retrieved chunks, tool results), not just the user's message. Scan it after rendering normalisation: strip zero-width and tag characters, decode HTML, extract text from images and documents, and inspect every modality the model receives. This is the only place AML.T0068, AML.T0129, AML.T0094 and AML.T0071 leave evidence.
- Guardrail verdicts as telemetry: send classifier blocks, near-misses and refusals to the SIEM keyed by identity and API key. Repeated refusals followed by success from the same key are the signature of jailbreak iteration.
- Output and rendering: alert on responses that contain external URLs, image references or citations to sources that were not in the retrieved set (AML.T0067, AML.T0067.000).
- Artifact intake: record hash, publisher, namespace and version for every model, dataset, package and MCP server; alert on namespace reuse, publisher changes and new versions that add network calls or new recipients (AML.T0074, AML.T0109, AML.T0111). Treat a model file that fails to deserialise in the scanner as suspicious, not as a scanner error (AML.T0076).
- Agent configuration and history integrity: monitor writes to agent configuration, rules files and memory, and keep server-side, append-only copies of conversation history so deletion or edits (AML.T0092, AML.T0138) cannot erase the record.
- Provider-side abuse: correlate requests across accounts by behaviour (prompt templates, capability targeting, timing) rather than metadata alone, because AML.T0140 and AML.T0141 are designed to defeat per-account and metadata-based views.
For telemetry design, see the AI Agent Observability and Incident Response Guide.
Mitigation priorities from ATLAS
Counting the ATLAS mitigation relationships that point at techniques in this tactic gives a clear picture of where the framework puts its weight, and where it has gaps.
| Mitigation | Mapped techniques in AML.TA0007 | Practical meaning |
|---|---|---|
| AML.M0020 Generative AI Guardrails | AML.T0054, AML.T0067, AML.T0068, AML.T0071, AML.T0094, AML.T0129 | Input and output filtering, layered and run on normalised, multimodal content |
| AML.M0035 AI Red Team | AML.T0015, AML.T0054, AML.T0068, AML.T0081 | Test guardrails and agents with obfuscated, delayed and multimodal payloads |
| AML.M0008 Validate AI Model | AML.T0015, AML.T0054, AML.T0129 | Robustness testing before and after deployment |
| AML.M0042 Threat Intelligence Program | AML.T0073, AML.T0107, AML.T0140 | Track impersonation campaigns, exploited flaws and abuse infrastructure |
| AML.M0014 Verify AI Artifacts and AML.M0023 Maintain an AI Bill of Materials | AML.T0109 | Pin, hash and inventory every AI component and its versions |
| AML.M0011 Restrict Library Loading and AML.M0025 Maintain AI Dataset Provenance | AML.T0074 | Control what can be loaded and where data came from |
| Predictive AI hardening set (AML.M0003, AML.M0006, AML.M0009, AML.M0010, AML.M0015) and AML.M0034 Deepfake Detection | AML.T0015 | Adversarial training, ensembles, sensor fusion, input restoration and adversarial input detection |
Eight techniques in the tactic have no ATLAS mitigation mapped: AML.T0076, AML.T0092, AML.T0097, AML.T0111, AML.T0123, AML.T0134, AML.T0138 and AML.T0141. For these, controls come from general security engineering: tamper-evident logging, sandbox design that does not reveal itself, scanner fail-closed behaviour, and provider-side correlation. The MITRE ATLAS Mitigations Guide covers each mitigation in depth.
Where identity fits
Evasion is far easier when the attacker already looks legitimate. Storm-2139 ran its jailbreak tooling through Azure OpenAI API keys that customers had exposed publicly, so to the provider the traffic came from paying customers. Distillation campaigns used fraudulent accounts and proxy services, which is exactly what Centralized AI Service Request Routing describes. Impersonation now includes posing as a trusted AI agent, not just a person. Practical consequences:
- Treat AI API keys as high-value secrets: scan for them, rotate them and bind them to expected networks and workloads. See the AI Agent Credential Lifecycle and Rotation Guide.
- Attribute every model call to an identity: a workload, agent or user, so guardrail verdicts can be aggregated per identity rather than per request.
- Make agent identity verifiable: agents that accept instructions from other agents should authenticate them, so a claimed persona is not enough.
Framework mapping
| Framework | Relevant item | Link to this tactic |
|---|---|---|
| MITRE ATT&CK v20 | TA0005 Stealth; TA0112 Defense Impairment | ATLAS cites TA0005 as its ATT&CK reference; impairment-style techniques such as AML.T0107 and AML.T0081 sit closer to TA0112 |
| OWASP Top 10 for LLM Applications 2025 | LLM01:2025 Prompt Injection | Obfuscation, jailbreaks, delayed and multimodal triggers |
| OWASP Top 10 for LLM Applications 2025 | LLM03:2025 Supply Chain | Masquerading, impersonation, rug pulls, reputation inflation, corrupt models |
| OWASP Top 10 for LLM Applications 2025 | LLM05:2025 Improper Output Handling | Output components manipulation and rendering paths used to exfiltrate |
| OWASP Top 10 for LLM Applications 2025 | LLM09:2025 Misinformation | Manipulated citations and false RAG entries presented as fact |
Common failure patterns
- One classifier, one modality: a single text injection filter in front of the model, with images, documents and hidden HTML passed through unchecked.
- Per-turn checks only: guardrails evaluate each message in isolation, so instructions planted in an earlier turn or retrieved later go unnoticed.
- Vetting once, trusting forever: MCP servers, packages and models are reviewed at adoption and then auto-updated, which is the window a rug pull uses.
- Scanner errors treated as clean: a model file that crashes the scanner is loaded anyway.
- History only on the client: chat history and agent logs live where the attacker or the agent can edit them, as the OpenAI and Hugging Face incident illustrates for agent-staged data.
- Alignment as the only gate: an agent with broad tool permissions relies on the model refusing, so a successful jailbreak becomes a successful action.
Coverage checklist for AML.TA0007
- Map every AI system to the techniques above and mark which detection and mitigation applies; leave no row blank without a recorded risk acceptance.
- Normalise and inspect the full assembled context, in every modality, before and after retrieval.
- Log guardrail verdicts with identity and key attribution, and alert on refusal-then-success patterns.
- Require human or policy approval for side-effecting tool calls, independent of the model's own judgement.
- Pin AI components by hash, review every version change, and monitor behaviour after updates.
- Configure model scanners to fail closed and to flag files that do not deserialise.
- Store conversation history, agent configuration changes and tool call logs server-side and append-only.
- Red team with obfuscated, delayed, multimodal and cloaked payloads at least before each major model or prompt change.
Practitioner checklist
- Inventory every AI-based security control you rely on and test it against AML.T0015 evasion.
- Add rendering normalisation (Unicode, HTML, OCR) in front of prompt injection classifiers.
- Strip or block external links and image references in model output unless allow-listed.
- Feed guardrail decisions into the SIEM with API key and identity.
- Scan code, tickets and documents for exposed AI API keys and rotate on discovery.
- Pin MCP servers and model versions, and review diffs before upgrading.
- Make conversation and agent logs tamper evident and retained outside the agent's reach.
- Schedule red team exercises against the six highest-ranked techniques in this hub.
Standards and references
- MITRE ATLAS: Defense Evasion (AML.TA0007)
- MITRE ATLAS: atlas-data repository (techniques, mitigations and case studies)
- MITRE ATLAS: Generative AI Guardrails (AML.M0020)
- MITRE ATT&CK: Stealth (TA0005)
- MITRE ATT&CK: Defense Impairment (TA0112)
- OWASP: LLM01:2025 Prompt Injection
- OWASP: LLM03:2025 Supply Chain
Related NHI Mgmt Group resources: MITRE ATLAS Guide, MITRE ATLAS Mitigations Guide, AI Guardrails and AI Firewalls Guide, Prompt Injection Guide, AI Supply Chain and AI-BOM Guide.