MAAIS — Multilayer Agentic AI Security Framework

A seven-layer defense-in-depth security framework for agentic AI systems, proposed in an arXiv preprint by Sunil Arora and John Hastings (arXiv 2512.18043v1, 2025-12-19). Methodologically grounded in Design Science Research (DSR) and validated via mapping the seven layers against MITRE ATLAS adversarial tactics. The framework’s most distinctive single contribution is CIAA — augmenting the classical CIA security triad (Confidentiality, Integrity, Availability) with Accountability as a first-class principle for agentic systems.

CIAA — augmented security triad

The classical CIA triad is well-established as a foundation of cybersecurity. MAAIS adds Accountability as a fourth principle specific to autonomous decision-making:

“In agentic AI, where decisions and actions may occur without direct human oversight, maintaining accountability is essential for security, transparency, and governance. Accountability enables identifying who is responsible for the AI Agent’s decisions and outcomes.”

The augmentation is a memorable shorthand for the wiki’s existing emphasis on action-to-identity tracing, audit chains of custody, and decision-rights documentation. CIAA is referenced inline in the CMM D1 (Governance & Accountability) as the foundation principle.

The seven layers

Each layer carries a defined set of controls; the layered architecture is explicitly framed as defense-in-depth + zero-trust.

#LayerRepresentative controls
1Infrastructure SecuritySecure hardware; secure compute; secure orchestration; supply-chain security; storage protection; network segmentation; secure CI/CD deployment
2Data SecurityData confidentiality; access control and data governance; differential privacy; data provenance and lineage; data integrity
3Model SecurityAdversarial defense techniques; model hardening and confidentiality; poisoning and backdoor detection; secure model deployment and execution
4Agent Execution and ControlExecution sandbox; policy (behavioral constraints) enforcement; runtime and safety verification; secure API and tools integration controls
5Accountability and TrustworthinessExplainable AI (XAI) techniques; bias detection and mitigation; AI system documentation; accountability and provenance mechanisms; human oversight and governance
6User and Access ManagementIdentity governance; access control (authorization); MFA and credential protection; privilege management and segregation of duties; continuous access monitoring and behavioral analytics
7Monitoring and AuditImmutable logging and audit trails; continuous monitoring and anomaly detection; agent behavioral analytics; threat intelligence and adaptive security policies; automated incident response mechanisms

MITRE ATLAS validation mapping

The paper’s validation step maps each ATLAS adversarial tactic to the MAAIS layer(s) responsible for mitigation:

MITRE ATLAS tacticMAAIS layer(s)
ReconnaissanceMonitoring and Audit
Initial AccessUser and Access Management; Infrastructure Security
ExecutionAgent Execution and Control
PersistenceAgent Execution and Control; Infrastructure Security
Privilege EscalationUser and Access Management; Infrastructure Security
Defense EvasionMonitoring and Audit; Model Security
Credential AccessUser and Access Management
DiscoveryMonitoring and Audit
CollectionData Security; Monitoring and Audit
Command and ControlAgent Execution and Control; Infrastructure Security
ExfiltrationData Security; Infrastructure Security
ImpactAccountability and Trustworthiness; Agent Execution and Control

The mapping is at the tactic level, not the technique level — the paper does not enumerate against the ~84 MITRE ATLAS techniques (AML.T####) individually. This is appropriate scope for a preprint introducing a framework but limits the validation to coverage rather than depth.

Cross-walk to wiki frameworks

MAAIS sits in conceptual space already occupied by several wiki pages. The crosswalk:

MAAIS LayerClosest wiki CMM domainClosest wiki RA planeCSA MAESTRO layer
1 InfrastructureD8 Supply Chain & Tooling (partial); cross-cutting infra(Hosting plane — implicit; not a named RA plane)L1 Foundation Models; L2 Data Operations
2 Data SecurityD6 Data, Memory & RAGData planeL2 Data Operations
3 Model SecurityD6 Data, Memory & RAG (partial); D4 Runtime & Guardrails (model-side)(Spans Runtime + Data planes)L1 Foundation Models
4 Agent Execution & ControlD3 Control & Least-Agency; D4 Runtime & GuardrailsControl plane; Runtime planeL4 Deployment & Infrastructure (sandboxing); L3 Agent Frameworks
5 Accountability & TrustworthinessD1 Governance & Accountability(Cross-cutting concern across planes)L7 Agent Ecosystem (audit/governance)
6 User & Access ManagementD2 Identity & AuthorizationIdentity plane(Cross-cutting)
7 Monitoring & AuditD7 Observability & Behavioral Monitoring; D1 Governance & AccountabilityObservability planeL5 Evaluation & Observability

MAAIS layers and CMM domains measure different things

MAAIS organizes by control surface (what kind of asset/process is being protected). The CMM organizes by operational domain with maturity tiers per domain (how systematically each surface is secured). The crosswalk above gives the closest content overlap; it does not imply equivalence. A control may show up under the same surface in both, with MAAIS describing what control to deploy and the CMM describing how mature that deployment needs to be at each maturity level.

Comparison with adjacent agentic-AI multi-layer frameworks

DimensionMAAIS (Arora & Hastings, 2025)CSA MAESTROWiki CMMAWS Scoping Matrix
Primary axisControl surface (7 layers)Threat model (7 layers)Operational domain × maturity (9 × 5+)Agency × autonomy (4 scopes)
ValidationMITRE ATLAS tactic mappingThreat enumeration per layerPer-domain Maps-to (multiple frameworks)Six security dimensions per scope
Maturity gatingNoneNone (gates from CSA ATF)L1–L5+ per domainNone (scope = deployment characterization)
PublicationarXiv preprint (Dec 2025)Standards body (CSA)This wiki (May 2026)Vendor blog (AWS, Nov 2025)
ScopeLayer-control contentLayer-threat contentMaturity progression + control contentScope characterization + control content

Stickiness assessment (2026-05-07, ~5 months post-publication)

  • CIAA (CIA + Accountability) framing: moderately sticky. The naming convention is memorable and easy to extend. Variants of CIA augmentation appear elsewhere (e.g., Parkerian Hexad: + Possession, Authenticity, Utility), so adding Accountability is a natural extension. Whether CIAA specifically (vs alternatives like CIA + Auditability or CIA + Authenticity) becomes the canonical agentic-AI augmentation is unclear at five months out. The wiki adopts CIAA as the inline anchor citation when discussing accountability as a security principle alongside CIA.
  • MAAIS name: likely not sticky. Multi-layer security frameworks for AI proliferate (CSA MAESTRO; the wiki’s CMM; AWS Scoping Matrix; OWASP). Each preprint that proposes one tends to use a distinctive acronym; absent significant practitioner traction or follow-on citation, MAAIS is unlikely to be remembered as the canonical seven-layer.
  • Seven-layer structure: convergent shape, distinct content. Multi-layer frameworks at 5–9 layers are common. MAAIS’s specific seven-layer breakdown overlaps substantially with MAESTRO’s seven layers but at different cuts. The convergence of “multi-layer ≈ 5-9 layers” is sticky; the specific layer breakdowns are bibliography-time choices.
  • MITRE ATLAS validation method: established practice. MAAIS uses tactic-level mapping (~12 tactics). The wiki’s CMM uses technique-level mapping (specific AML.T#### IDs). Both methods are sound; MAAIS’s choice is appropriate for its scope but limits depth.
  • Design Science Research methodology: academic-only. DSR is a standard IS research framing; doesn’t propagate into practitioner discourse.

Limitations

  • Tactic-level validation, not technique-level. The MITRE ATLAS mapping covers 12 tactics but does not address specific AML.T#### techniques. For deeper validation, pair with the wiki’s CMM domain Maps-to lines that cite specific techniques.
  • No threat enumeration per layer. Unlike MAESTRO which threat-models each layer explicitly, MAAIS lists layer controls without specifying which threats motivate each control.
  • No prompt-injection / lethal-trifecta treatment. The paper does not name the Lethal Trifecta, indirect prompt injection, MCP-specific risks, agent-to-agent (A2A) orchestration risks, or promptware. These are the load-bearing threat classes in current practitioner discourse but absent from this framework.
  • No agency-vs-autonomy distinction. Uses both terms but does not formalize the split (now anchored from the AWS Scoping Matrix).
  • No identity / NHI deep treatment. Layer 6 (User & Access Management) lists controls but does not address the NHI lifecycle, identity-credential coupling, or credential proxy pattern that the wiki treats as load-bearing.

Concepts MAAIS surfaced that the wiki had under-treated

A May 2026 audit of MAAIS against the wiki identified four real coverage gaps that the wiki’s threat-modeling had under-prioritized in favor of orchestration-layer (prompt-injection-class) threats. New concept pages now anchor each:

  • Differential Privacy — was zero-coverage; now the canonical defense citation for inversion and membership-inference attacks; control surfaces in CMM D6 Maps-to.
  • Model-Layer Attacks — extraction / inversion / membership-inference were under-treated; the new page enumerates them as a class with shared defensive primitives; cites MITRE ATLAS techniques (AML.T0024, AML.T0044, AML.T0048).
  • Agent Availability Threats — runaway / recursion / resource-exhaustion as a named class. The Lethal Trifecta is a C+I model; availability is a separate axis that the MAAIS CIAA framing surfaces and this concept page anchors.
  • Operational XAI for Action Gating — distinct from mechanistic interpretability; covers runtime-gated justifications for high-impact actions. MAAIS Layer 5 named “XAI Techniques” without specifying the operational vs research split; the wiki now documents both.

Use cases

  • Lecture / course-material reference — DSR-based academic paper appropriate for graduate-level AI security curricula. The seven-layer breakdown is teachable in a single lecture.
  • CIAA citation anchor — when discussing accountability as a security principle alongside CIA, cite this paper’s section IV-A as the published source of the augmentation.
  • MITRE ATLAS coverage check — the tactic-level mapping table is a useful starting point if validating that a security architecture covers all twelve ATLAS tactics. Pair with technique-level Maps-to in the wiki’s CMM for the depth dimension.

Provenance

Authored by Sunil Arora and John Hastings; arXiv preprint 2512.18043v1 submitted 2025-12-19; arXiv categories cs.CR / cs.AI / cs.CY; ACM classes K.6.5 (Computing Milieux — Management of Computing and Information Systems / Security and Protection), D.4.6 (Operating Systems — Security and Protection), I.2.11 (Artificial Intelligence — Distributed Artificial Intelligence). DOI: 10.48550/arXiv.2512.18043. No journal reference (preprint). Source-summary at the paper page.