AxiLayer AI crestAxiLayerAI™
AxiLayer AI · Frameworks

MITRE ATLAS: attacks on AI, catalogued

The adversary tactics against AI systems and agents, mapped to the controls that block them and the tests that prove it.

Run the Agent Exposure CheckSee the mapping

Last verified: 30 October 2026

Home / Frameworks / MITRE ATLAS
The framework

The vocabulary of the AI red team

What it is
MITRE ATLAS, the Adversarial Threat Landscape for AI Systems: a knowledge base of tactics, techniques and case studies for attacks on AI, built on the ATT&CK model and extended through 2025 and 2026 with techniques against agents, tools and retrieval.
Why it matters
It is the language red teams, insurers and regulators use to describe an AI attack. EU AI Act Article 15 and ISO/IEC 27090 ask for resistance to the same attacks ATLAS catalogues.
How AxiLayer uses it
The authorized adversarial tests in the agentic test suite are built from ATLAS techniques; findings and incidents are tagged with them so a CISO can read the record in the vocabulary of the SOC.
Mapping

Each requirement, the control that meets it

Every row is a control that is live in AxiSentinel or AxiReady today, with the phase it belongs to.

RequirementWhat AxiSentinel and AxiReady doPhases
Reconnaissance and resource developmentAI radar shows what is exposed; AI-BOM and OSV.dev lookup for known weaknesses in components.01, 09
Initial access and ML model accessLLM proxy and MCP Gateway are the only paths to models and tools; unapproved servers blocked; per-agent identity.03, 04, 05
Execution and persistenceCode-execution tools gated; behavior guard flags an agent that keeps acting outside its normal; quarantine.03, 07, 08
Privilege escalation and credential accessNamed identity, role scope and credential reference on every action; shared or over-broad credentials surface as findings.04, 10
Defense evasionEvery call logged in a tamper-evident record; gateway blocks recorded; re-tests after every change.09, 10
Discovery and collectionData scope per agent enforced at the gateway; unusual access flagged.04, 07
ML attack staging (poisoning, adversarial examples, prompt injection)Adversarial and injection cases in the authorized test suite; drift detection; re-test on data change.07, 10
ExfiltrationEgress through the gateway only; destination anomalies flagged; stop on one confirmed decision.05, 07, 08
ImpactHigh-impact actions held for a person; kill switch enforced at ingest and proxy; incident clock and evidence preserved.06, 08, 09

Sources: MITRE ATLAS knowledge base and case studies; EU AI Act Article 15; ISO/IEC 27090:2026.

Evidence

What the Compliance Passport carries for this framework

Adversarial test report
Which ATLAS techniques were exercised against each AI system, with the result, signed.
Scorecard by tactic
Events seen, blocked or held, missed, grouped by ATLAS tactic.
Incident record
Technique tags on each incident, with the Article 73 clock and the evidence chain.
Three layers, twelve phases
Twelve phases. One platform. One signature.
Phase 01FindDiscover and score
Phase 02FindClassify and plan
Phase 03ControlApprove every tool
Phase 04ControlName every agent
Phase 05ControlGate every action
Phase 06ControlHold for a person
Phase 07ControlWatch behavior
Phase 08ControlQuarantine and stop
Phase 09ControlMonitor to the article
Phase 10ProveTest and re-test
Phase 11ProveSign the evidence
Phase 12ProveVerify, free, forever

AxiReady finds. AxiSentinel gates, holds, quarantines and stops, Always On or On Demand. A qualified AxiLayer AI auditor signs against MITRE ATLAS.

Next step

See where your agents stand in three minutes

Twelve questions, three minutes, a band and three first steps. Then the 48-Hour Readiness Report for your estate.

Run the Agent Exposure CheckThe Agent Control PlaneSchedule a Scoping Call