Skip to content

AI Security Infrastructure Controls

The infrastructure layer that makes behavioural AI security enforceable.

Companion to AI Runtime Behaviour Security, which defines the three-layer behavioural security pattern (Guardrails → LLM-as-Judge → Human Oversight). This repo defines the 80 technical infrastructure controls that underpin that pattern.

License: MIT


The Problem

You can't enforce behavioural security controls on infrastructure you don't control. The parent framework tells you what to enforce. This repo tells you how to enforce it:


Architecture

How Technical Controls Enable the Three-Layer Pattern

Relationship to Parent Framework

IAM Control Architecture

IAM Control Layers

Logging Pipeline and Detection Architecture

Logging Architecture

Network Zone Architecture

Network Zones


Control Domains

Identity & Access Management — 8 controls

Authentication, least privilege, control plane separation, agent tool constraints, human approval workflows, session credentials, credential exposure prevention, and access auditing.

Control Purpose Risk Tiers
IAM-01 Authenticate all entities All
IAM-02 Enforce least privilege All
IAM-03 Separate control/data planes Tier 2+
IAM-04 Constrain agent tool invocation Tier 2+ (agentic)
IAM-05 Human approval for high-impact actions Tier 3+ (agentic)
IAM-06 Session-scoped credentials Tier 2+
IAM-07 Prevent credential exposure in context All
IAM-08 Audit all access changes Tier 2+

Logging & Observability — 10 controls

Model I/O logging, guardrail decisions, Judge evaluations, agent chains, drift detection, injection detection, log integrity, retention, redaction, and SIEM correlation.

Control Purpose Risk Tiers
LOG-01 Log all model inputs and outputs All
LOG-02 Log guardrail decisions All
LOG-03 Log Judge evaluations Tier 2+
LOG-04 Log agent decision chains Tier 2+ (agentic)
LOG-05 Detect behavioural drift Tier 2+
LOG-06 Detect prompt injection attempts All
LOG-07 Protect log integrity Tier 2+
LOG-08 Enforce retention policies All
LOG-09 Redact sensitive data in logs All
LOG-10 Correlate with enterprise SIEM Tier 3+

Network & Segmentation — 8 controls

Six-zone architecture, guardrail bypass prevention, Judge isolation, agent egress controls, ingestion/runtime separation, control plane protection, API gateway enforcement, and cross-zone monitoring.

Control Purpose Risk Tiers
NET-01 Define network zone architecture All
NET-02 Prevent guardrail bypass at network level All
NET-03 Isolate Judge evaluation infrastructure Tier 2+
NET-04 Control agent egress destinations Tier 2+ (agentic)
NET-05 Separate ingestion from runtime Tier 2+
NET-06 Protect control plane network path All
NET-07 Enforce API gateway as single entry All
NET-08 Monitor cross-zone traffic Tier 2+

Data Protection — 8 controls

Data classification, minimisation, PII detection/redaction, access-controlled RAG, encryption, response leakage prevention, conversation history management, and evaluation data protection.

Control Purpose Risk Tiers
DAT-01 Classify data at AI boundaries All
DAT-02 Enforce data minimisation All
DAT-03 Detect and redact PII in model I/O All
DAT-04 Enforce access-controlled RAG retrieval Tier 2+
DAT-05 Encrypt data at rest and in transit All
DAT-06 Prevent sensitive data leakage via responses Tier 2+
DAT-07 Manage conversation history retention All
DAT-08 Protect evaluation data sent to Judge Tier 2+

Secrets & Credentials — 8 controls

Context window isolation, short-lived tokens, centralised vault, credential scanning, rotation on exposure, agent credential isolation, endpoint protection, and code scanning.

Control Purpose Risk Tiers
SEC-01 Never inject credentials into context windows All
SEC-02 Use short-lived, scoped tokens All
SEC-03 Centralise secrets in a vault All
SEC-04 Scan model I/O for credential patterns All
SEC-05 Rotate credentials on exposure All
SEC-06 Isolate agent credentials per session Tier 2+ (agentic)
SEC-07 Protect model endpoint credentials All
SEC-08 Scan code and config for embedded credentials All

Supply Chain — 8 controls

Model provenance, risk assessment, RAG data integrity, fine-tuning security, tool supply chain, safety model integrity, AI-BOM, and vulnerability monitoring.

Control Purpose Risk Tiers
SUP-01 Verify model provenance and integrity All
SUP-02 Assess model risk before adoption All
SUP-03 Verify RAG data source integrity Tier 2+
SUP-04 Secure fine-tuning pipeline Tier 2+
SUP-05 Audit tool and plugin supply chain Tier 2+ (agentic)
SUP-06 Verify guardrail and safety model integrity All
SUP-07 Maintain AI component inventory (AI-BOM) All
SUP-08 Monitor for model and dependency vulnerabilities All

Incident Response — 8 controls

AI-specific incident categories, detection triggers, containment procedures, rollback capability, investigation for non-deterministic systems, communication protocols, post-incident review, and enterprise IR integration.

Control Purpose Risk Tiers
IR-01 Define AI-specific incident categories All
IR-02 Establish detection triggers from logging All
IR-03 Define containment procedures All
IR-04 Implement model rollback and guardrail hot-reload Tier 2+
IR-05 Define investigation for non-deterministic systems Tier 2+
IR-06 Establish communication protocols Tier 2+
IR-07 Conduct AI-specific post-incident review All
IR-08 Integrate with enterprise IR processes All

Agentic AI Controls

Additional controls for systems where AI agents invoke tools, generate code, delegate to other agents, or take autonomous actions.

Tool Access Controls — 6 controls

Control Purpose
TOOL-01 Declare tool permissions explicitly before deployment
TOOL-02 Enforce permissions at the gateway, not at the agent
TOOL-03 Constrain tool parameters to declared bounds
TOOL-04 Classify tool actions by reversibility and impact
TOOL-05 Rate-limit tool invocations per agent and per tool
TOOL-06 Log every tool invocation with full context

Session & Scope — 5 controls

Control Purpose
SESS-01 Define and enforce session boundaries with automatic expiry
SESS-02 Isolate sessions from each other
SESS-03 Limit session scope to a declared task
SESS-04 Implement progressive trust within sessions
SESS-05 Clean up session state on termination

Delegation Chains — 5 controls

Control Purpose
DEL-01 Enforce least delegation (no privilege escalation)
DEL-02 Maintain complete audit trail across chains
DEL-03 Limit delegation depth
DEL-04 Require explicit delegation authorisation
DEL-05 Propagate user identity through chains

Sandbox Patterns — 6 controls

Control Purpose
SAND-01 Execute generated code in isolated sandboxes
SAND-02 Restrict sandbox file system access
SAND-03 Restrict sandbox network access
SAND-04 Enforce resource limits on execution
SAND-05 Prevent persistent state escaping sessions
SAND-06 Scan generated code before execution

Regulatory & Standards Mappings

Every control maps to the three-layer behavioural model and to established standards:

Mapping Coverage
Controls → Three Layers All 80 controls × Guardrails, Judge, and Human Oversight
ISO 42001 Annex A Full Annex A technical infrastructure mapping
NIST AI RMF 51 subcategories across Govern, Map, Measure, Manage
OWASP LLM Top 10 All 10 risks + OWASP Agentic Top 10

Platform Implementation Patterns

Reference patterns for implementing these controls on specific platforms:

Platform Covers
AWS Bedrock Bedrock Guardrails, Knowledge Bases, Agent Framework, VPC architecture, IAM roles
Azure AI Azure OpenAI, AI Content Safety, APIM, Entra ID PIM, Sentinel
Databricks Mosaic AI Gateway, Model Serving, Unity Catalog, Vector Search, Inference Tables

Diagrams

Fourteen SVG diagrams covering IAM architecture, logging pipelines, network zones, agent egress flows, credential isolation, data classification, incident classification, delegation permission models, and more. All in diagrams/.


Design Principles

  1. Vendor-neutral. Controls are defined in terms of what must be enforced, not which product enforces it. Platform-specific guidance lives in reference/platform-patterns/.
  2. Three-layer aligned. Every control maps back to how it supports Guardrails, Judge, or Human Oversight.
  3. Risk-tiered. Controls specify which risk tiers they apply to, enabling proportionate implementation.
  4. Practitioner-focused. Written for security architects and engineers, not executive summaries.
  5. Agentic-aware. Agent-specific controls are treated as first-class concerns, not afterthoughts.
  6. Infrastructure beats instructions. Security controls are enforced via deterministic infrastructure (gateways, network policy, vaults), never via prompt instructions that can be overridden.

Repo Structure

See REPO-STRUCTURE.md for the full layout and rationale.


Contributing

Feedback, corrections, and extensions welcome. This is a living framework.


Author

Jonathan Gill — Cybersecurity practitioner with 30+ years in IT and 20+ years in enterprise cybersecurity, currently focused on AI security governance.


AI Runtime Behaviour Security, 2026 (Jonathan Gill).