Skip to content

Controls

This is where you choose the controls. Once you have threat modelled the system and sized its risk tier, this section gives you the control set to apply: the three-layer pattern, agentic and IAM controls, the reviewing layer (scanners, semantic firewall, model-as-judge), and the specialised controls particular deployments need. Select what fits the risk, adapt it to how your organisation works, and consciously deselect the rest.

Canonical home: airuntimesecurity.io

The runtime control definitions are maintained in depth on AI Runtime Security. These pages mirror that material so an architect can design the control set here and hand off to AIRS for runtime operation. Where the two differ, airuntimesecurity.io is authoritative.

Reading Order

Start with the essentials, then branch into specialised topics based on your deployment:

Essential (read in order): 1. Risk Tiers - classify your system 2. Risk Assessment - quantify control effectiveness and residual risk per tier 3. Controls - implement the three-layer pattern 4. Agentic - add controls if your agent has tool access 5. IAM Governance - identity, lifecycle, delegation 6. Judge Assurance - measure and calibrate the model-as-judge within the reviewing layer 7. Checklist - track implementation progress

Specialised (read based on your deployment type):

If you're deploying... Read
Multimodal models (image, audio, video) Multimodal Controls
Reasoning models (chain-of-thought) Reasoning Model Controls
Streaming responses Streaming Controls
Persistent memory or long context Memory and Context
Multi-agent systems Multi-Agent Controls then MASO (Multi-Agent Security Operations)
Open-weight / self-hosted models Open-Weight Models Shift the Burden

PACE resilience (read after controls): - Control Layer Resilience - PACE for each control layer - PACE for Agentic AI - PACE for agentic deployments - PACE Checklist - verify your fail postures

The Fundamental Shift

Traditional software can be tested before deployment. AI cannot - not fully.

Traditional Software AI Systems
Deterministic outputs Non-deterministic
Testable at design time Emergent behavior
Known failure modes Adversarial discovery

The shift: From design-time assurance to runtime behavioral monitoring.

The Pattern

The industry is converging on three layers:

Layer Function Timing
Guardrails Block known-bad inputs/outputs deterministically Real-time
Reviewing controls Second opinion on unknown-bad: scanners, semantic firewall, model-as-judge Inline and async
Human Oversight Decide, act, remain accountable As needed

Guardrails prevent. Reviewing controls catch what they miss. Humans decide. Circuit breakers contain. The judge within the reviewing layer informs the decision; it never replaces the deterministic guardrails beneath it. The circuit breaker sits behind all three as containment, not a behavioural layer, tripping to a non-AI fallback when they are bypassed or overwhelmed, and maps to PACE Emergency.

Where This Pattern Exists

This isn't theoretical. Production implementations include:

Platform Implementation
NVIDIA NeMo Guardrails Input, dialog, retrieval, execution, output rails
LangChain Middleware + human-in-the-loop
Guardrails AI Open-source validator framework
Galileo Eval-to-guardrail lifecycle
DeepEval Model-as-Judge evaluation
AWS Bedrock Guardrails Managed filtering
Azure AI Content Safety Content moderation

What has been missing: clear guidance on why this pattern is necessary and how to implement it proportionate to risk, in a way that respects how each organisation actually works.

Scope

In: Custom LLM apps, AI decision support, document processing, agentic systems
Out: Vendor AI products, model training, data preparation

Quick Start

1. Classify Your System

Tier Profile Examples
CRITICAL Direct decisions, customer/financial/safety impact Credit decisions, fraud blocking
HIGH Significant influence, sensitive data Customer service with account access
MEDIUM Moderate impact, human review expected Internal Q&A, document drafting
LOW Minimal impact, non-sensitive Public FAQ, suggestions

2. Select Controls

Control LOW MEDIUM HIGH CRITICAL
Input guardrails Basic Standard Enhanced Maximum
Output guardrails Basic Standard Enhanced Maximum
Reviewing controls Scanners Scanners + sampled judge + Semantic firewall, inline judge Full layer, inline on high-impact
Human review Exceptions Sampling Risk-based All significant

3. Implement in Order

  1. Logging - Can't evaluate what you don't capture
  2. Basic guardrails - Block obvious attacks
  3. Reviewing controls in shadow mode - Scan and evaluate without action
  4. HITL queues - Somewhere for findings to go
  5. Operationalise - Act on findings, tune continuously

Core Documents

Document Purpose
Risk Tiers Classification criteria, control mapping
Risk Assessment Quantitative control effectiveness, residual risk analysis, worked examples per tier
Controls Guardrails, reviewing controls, HITL implementation
Agentic Additional controls for agents
IAM Governance Identity governance, agent lifecycle, delegation, threats
Checklist Implementation tracking
Emerging Controls Multimodal, reasoning, streaming overview

Specialized Controls

Document Purpose
Judge Assurance Judge accuracy measurement and calibration
Multi-Agent Controls Controls for multi-agent systems
Multimodal Controls Controls for image, audio, and video AI
Memory and Context Long context and persistent memory controls
Reasoning Model Controls Controls for chain-of-thought reasoning models
Streaming Controls Controls for real-time streaming outputs

PACE Sections

Document Purpose
PACE Controls Section PACE framework - controls
PACE Agentic Section PACE framework - agentic controls
PACE Checklist Section PACE framework - implementation checklist

Architecture Overview

Architecture Overview

Key Principles

  1. Match controls to risk - Apply the right controls at the right time for the right purposes. Do not over-engineer low-risk systems
  2. Respect organisational context - Every organisation has its own structures and ways of working. Select and adapt controls accordingly
  3. Guardrails are necessary but not sufficient - They miss novel attacks and nuance
  4. The reviewing layer is assurance, not a hard gate - Scanners and the semantic firewall add deterministic and semantic checks; the model-as-judge informs the decision but is probabilistic and can be fooled, so it never replaces the deterministic guardrails beneath it
  5. Infrastructure beats instructions - Enforce technically, not via prompts
  6. Assume bypasses happen - Design for detection, not just prevention
  7. Humans remain accountable - AI assists; humans own outcomes