RESEARCH FRAMEWORK · GAP ANALYSIS

A control-maturity instrument for organizations introducing LLM-enabled systems.

This library documents the research design, evaluation instrument, and analytical artifacts for an study of how mature corporate information security architectures hold up against the structurally different exposure interface introduced by Large Language Models. The framework couples ISO 27002:2022 controls with an adapted C2M2 v2.1 maturity scale, contextualized across four LLM deployment archetypes.

Browse the library Start with the research design Event lifecycles Deployment archetypes Critical evaluation 1009 LTD baseline
RESEARCH CONTEXT

The central question is structural, not technical.

Assuming an organization possesses mature corporate governance — aligned with ISO 37004 — are its existing organizational and technical information security controls (aligned with ISO 27002:2022) objectively mature enough to govern the introduction of LLM-enabled systems? The framework isolates control maturity from governance maturity, then tests each applicable control against four deployment archetypes and eight risk domains. Scoring uses an adapted C2M2 instrument; aggregation produces an objective Priority Quota and a qualitative gap typology.

8
Top-level LLM risk domains, derived from the AI Act and recent attack surveys
4
Deployment archetypes: Direct, Enterprise, Application, Fine-Tuned
71
ISO 27002:2022 organizational and technical controls in scope
4
Maturity Indicator Levels (0–3), adapted from C2M2 v2.1
METHODOLOGICAL ANCHOR

1009 LTD — a simulated, fully documented SME baseline.

To evaluate control sufficiency under a realistic but reproducible profile, the framework uses 1009 LTD — a simulated small and medium-sized enterprise modelled as 40 employees across four divisions, with twelve documented internal policies spanning Information Security, Risk Management, Business Continuity, Identity Management, and Physical Security. Anchoring the analysis to a published profile solves two problems at once: it bypasses confidentiality constraints that limit empirical AI security work, and it gives the methodology a reproducible reference scenario.

Methodological note

Because the study assumes a baseline of mature corporate governance, the C2M2 instrument is methodologically "bent" away from organizational governance attributes (such as executive sponsorship or enterprise policy formulation, which are assumed to be present) and toward the operational, architectural, and technical resilience of specific ISO 27002 controls under LLM conditions.

LIBRARY · 7 + 1 DOCUMENTS

Each stage of the framework is a standalone page.

Documents follow the analytical flow from research design through risk scenarios, controls, maturity, procedure, and gap typology. A critical evaluation is provided as a separate page.

DOC 00 · MODULE
Research Design
Exploratory qualitative gap analysis. Isolates control maturity from governance maturity via an adapted C2M2 v2.1 instrument.
MethodologyC2M2
Read module
DOC 01 · MODULE
Deployment Archetypes
Four ways organizations adopt LLMs: Direct external use, Enterprise-embedded, Application-level integration, and Fine-tuned/internally managed.
DirectEnterpriseApplicationFine-Tuned
Read module
DOC 02 · MODULE
LLM Risk Scenarios
Eight top-level risk domains, expressed as Cause, Event, Consequence. Prompt injection and shadow AI receive special treatment.
8 domains
Read module
DOC 03 · MODULE
Control Baseline
ISO 27002:2022 organizational and technical controls. Scope rationale, omitted control families, and rationale for the GAISO link.
ISO 27002
Read module
DOC 04 · MODULE
Maturity Model
Adapted C2M2 Maturity Indicator Levels (MIL 0–3). Approach and Management progression measured via four qualitative criteria.
MIL 0–3
Read module
DOC 05 · MODULE
Gap Analysis Procedure
Three stages: contextualize the risk per archetype, map controls (Direct/Indirect/Fails), assess maturity against LLM context.
3 stages
Read module
DOC 06 · MODULE
Priority Quota & Gap Types
Quantitative Priority Quota (Urgency × Capability Gap) feeding a four-band heatmap. Four qualitative gap types: Design, Applicability, Visibility, Accountability.
Heatmap4 gap types
Read module
DOC 07 · CRITICAL EVALUATION
Critical Evaluation
An assessment of the research design (research design module) against its companion gap analysis module and the underlying scoring instrument. Strengths, structural inconsistencies, and recommendations.
Peer reviewOpen issues
Read evaluation
EVENT LIFECYCLES · DOWNLOADABLE TAXONOMY

Eight LLM threat domains, each as a full kill-chain taxonomy.

Every risk domain identified in the framework is documented as a standardized event lifecycle: phase taxonomy (Reconnaissance, Weaponization, Exploitation, C2, Actions on Objectives), TTP mapping against MITRE ATLAS / ATT&CK, preventive/detective/corrective control intersections, the telemetry needed for audit-grade evidence, and the auditor's CCCE perspective (Criteria · Condition · Cause · Effect · Recommendation). Each file is a plain-text master context document used as the methodological backbone for Step 2 of the gap analysis.

RS1 · DATA POISONING
Manipulate the Training Data Set
Supply-chain attack on the training corpus. Adversary submits crafted samples upstream (open datasets, feedback loops, data lakes) so a routine training run installs a backdoor in the model's weights. Triggered later via specific inputs.
RS1MITRE ATLAS T0020DVC / Golden Dataset
Download .txt
RS2 · MODEL POISONING
Manipulate Pre-Trained Components
Direct supply-chain compromise of the foundational model. Two flavors: semantic weight manipulation surviving fine-tuning, and code execution via unsafe serialization (pickle / torch.load). Demands a walled-garden artifact registry.
RS2ATLAS T0010.safetensors only
Download .txt
RS3 · ADVERSARIAL EXAMPLES
Inputs Engineered to Cause Mistakes
Inference-phase attack. Mathematically perturbed inputs (FGSM, PGD) force the model across a decision boundary while remaining imperceptible to humans. Defeated by adversarial training, feature squeezing and OOD detection.
RS3ATLAS T0043 / T0015OOD detection
Download .txt
RS4 · MODEL EVASION
Semantic Bypass of AI-Driven Controls
Tactical evasion of AI classifiers, safety guardrails, and AI-AV through jailbreaking, obfuscation, or structural manipulation. Requires independent input/output guardrails (LlamaGuard, NeMo) and deterministic fallback controls.
RS4ATLAS T0015 / T0054Prompt Shields
Download .txt
RS5 · CONFIDENTIALITY ATTACKS
Data Extraction, Inversion, Model Theft
Exploits the model's tendency to memorize. Adversary harvests outputs to reconstruct PII, secrets, or full model weights. Mitigated by training-set deduplication, differential privacy, and DLP scanning on model outputs.
RS5ATLAS T0024 / T0044DP-SGD
Download .txt
RS6 · MODEL FLAWS
Inherent Defects, Hallucination, DoS
Exploits native architectural limits: tokenizer bugs, sponge examples (resource exhaustion), alignment gaps, and confident hallucination. Defeated by AI red-teaming, hard computational bounds, and grounded output verification.
RS6ATLAS T0029 / T0031Circuit breakers
Download .txt
RS7 · PROMPT INJECTION
Direct and Indirect Instruction Hijack
Architectural flaw: LLMs cannot reliably distinguish instructions from data. Direct (jailbreak) and indirect (payload embedded in RAG content or web pages) variants. Enables excessive agency, exfiltration via forced Markdown rendering, RCE-by-proxy.
RS7OWASP LLM01HITL required
Download .txt
RS8 · SHADOW AI
Unauthorized Employee Experimentation
Authorized employees acting outside governance. Unmonitored SaaS provisioning, OAuth grants to AI apps over corporate data, local LM runtimes, and unvetted code ingestion. Demands CASB, endpoint DLP, application whitelisting, and a sanctioned internal AI gateway.
RS8CASB / DLPInsider profile
Download .txt
LLM DEPLOYMENT ARCHETYPES · MASTER CONTEXT FILES

Four ways an SME adopts LLMs — each with its own threat surface.

Each archetype is published as a downloadable master context file used to contextualize every ISO 27002 control during Step 2 of the gap analysis. The files capture the organizational baseline (1009 LTD as CASP), the operational modality, primary threat vectors, and the applicable governance criteria policy-by-policy. They are the input the framework feeds into the C2M2-adapted maturity instrument when scoring control sufficiency.

ARCHETYPE A · DIRECT
Direct External LLM Use
Employees use public LLM web interfaces (ChatGPT, Claude, Gemini) without enterprise contracts. Primary failure modes: uncontrolled prompting and data disclosure, weak organizational visibility (Shadow AI), and bypass of DLP / web-filtering where the "deny all" default is defective.
A · DirectP08 · ISPP06 · OpsMgmtP11 · NetSec
Download .txt
ARCHETYPE B · ENTERPRISE
Enterprise-Embedded Assistants
M365 Copilot, Google Workspace Gemini, Slack AI — assistants embedded in the productivity suite and indexing the whole tenant via Microsoft Graph and equivalents. Failure modes: weaponization of over-permissioning, normalization of Shadow AI workflows, and contextual leakage of confidential data through AI-generated outputs.
B · EnterpriseP01 · ACP12 · IdMP09 · RM
Download .txt
ARCHETYPE C · APPLICATION
Application-Level Integration (External APIs)
Custom applications integrating external LLM APIs (chatbots, code-review assistants, RAG pipelines, agentic workflows). Failure modes: prompt injection & improper output handling, RAG/vector-database poisoning, and excessive agency granted to AI service accounts on downstream systems.
C · ApplicationP11 · NetSecP07 · IncMgmtExcessive agency
Download .txt
ARCHETYPE D · FINE-TUNED
Internally Managed / Fine-Tuned Models
Open-weight or proprietary models hosted in-house and fine-tuned on internal datasets. Demands mature MLOps. Failure modes: training-data poisoning & memorization of CASP secrets, theft of fine-tuned model weights, and capacity / performance collapse of GPU infrastructure not covered by traditional IT metrics.
D · Fine-TunedP04 · AssetP02 · BackupP10 · PhySec
Download .txt
GAP ANALYSIS v4 · PRIORITY OUTPUT

What the updated scoring instrument actually returned.

The v4 (atnaujintas) workbook defines a nominal 71-control Organizational + Technological scope and actively scores 70 controls against the 32 cross-mapped risk scenarios (8 domains × 4 archetypes). For each control the instrument computes Direct / Indirect mapping counts, a priority score (Urgency × applicability), an adapted C2M2 maturity rating across four criteria (Coverage, Specificity, Operationalizability, Adaptability), the resulting capability gap against the LLM-specific target maturity, and a final Priority Quota — bucketed into Critical / High / Medium bands for management triage.

70
Active controls scored end-to-end (71 in nominal Organizational + Technological scope)
6
Critical-priority controls (PQ = 20) demanding immediate intervention
31
High-priority controls (PQ = 15–16) requiring near-term capability lift
33
Medium-priority controls (PQ = 12) for the second planning horizon
IDControlUrgencyMaturityTargetCapability gapPriority QuotaBand
O24Information security incident management planning and preparation515420Critical
O25Assessment and decision on information security events515420Critical
O26Response to information security incidents515420Critical
O27Learning from information security incidents515420Critical
O28Collection of evidence515420Critical
O21Managing information security in the ICT supply chain515420Critical
O15Access control415416High
O31Legal, statutory, regulatory and contractual requirements415416High
T09Configuration management415416High
T15Logging525315High
T16Monitoring activities525315High

Excerpt from the GAP Analysis v4 (atnaujintas) workbook · 6 Critical + 5 of 31 High-band controls shown. The full Step 1 / Step 2 / Step 3 / Heatmap workbook is the analytical artifact behind these numbers.

Headline finding

The Critical band is entirely populated by incident management and supply-chain controls (O21, O24–O28). The organization's ability to detect, decide on, respond to, learn from, and evidence LLM-related events is the structural weakness — not its preventive controls. This is consistent with the Shadow AI, Prompt Injection and Confidentiality Attack lifecycles, which all collapse traditional incident detection because the malicious activity looks like legitimate HTTPS / API traffic.

Open the interactive GAP Console Priority Quota methodology
COVERAGE MATRIX

How risk domains map to control families.

Each of the eight LLM risk domains is initially mapped against ten generic control domains (D1–D10), which group ISO 27002:2022 clauses by functional purpose. The dot indicates the primary functional mapping; secondary and indirect mappings are evaluated separately in the gap analysis procedure.

Risk domainD1
Gov & Policy
D2
Data
D3
IAM
D4
SecDev
D5
Supplier
D6
Logging
D7
Training
D8
Incident
D9
Change
D10
Assurance
Data poisoning·········
Model poisoning·········
Adversarial examples·········
Model evasion·········
Confidentiality attacks········
Model flaws·········
Prompt injection·······
Shadow AI·······

● Primary functional mapping. Indirect mappings are evaluated in Stage 2 of the procedure; full applicability under each archetype is decided in Stage 3 (see Gap analysis procedure).