AI Harness
AI Harness Capability Matrix: 18 Governed Components for Accountable AI
© 2026 Richard E. Davis · davisAIgov AI Harness Capability Matrix. Eighteen governed components surround one AI model: 19 nodes in the complete reference diagram. Conceptual framework; not a claim that controls are deployed.
Your AI model operates inside a harness of connected components that provide data, memory, tools, credentials and authority. Each connection adds an attack surface. Check which systems a component can reach, what it can change, who can approve the action and whether the control actually stopped an unauthorized request.
The model generates a proposed answer or action. The harness supplies the data, memory, tools, credentials and connections needed to carry it out. If a retrieved document gives malicious instructions or a tool has excessive permissions, risk can move beyond the model. Independent authorization checks, controlled execution and separate audit records help keep business decisions within approved limits.
The operating components
The 19 components include the model, the systems it connects to and the oversight functions around them. For each component, ask: What can it read or change? Which attack would matter? Where is the independent stop? What test or log confirms it worked?
Orchestrator / agent loop
RAG subsystem
Memory & state
Context assembly
Input / output guardrails
Policy / authorization gate
Credential broker
Extensions / MCP / skills
Tool dispatcher
Sandbox / downstream systems
Observability & audit
Generative model
Identity & delegation
Evaluation & conformance
Human oversight console
AI BOM & attestation
Authority & consumption
Transparency & recourse
Inference is not authority.
A model can suggest a payment, an account change or a disclosure. The business must authorize that exact request before a connected tool performs it. A fluent response or a successful login is not permission to act.
AI may recommend an action. Your business rules decide whether it happens. Before AI sends customer data, issues a payment or changes a live record, a separate permission check must allow it, hold it for a named reviewer or block it. This is the policy / authorization gate: a control outside the AI, not an instruction asking it to behave.
For example, AI may prepare a refund, but a payment above your approved limit must wait for the designated reviewer. The payment system checks that permission matches the amount and recipient before paying. Keep records of the request, the decision and what actually happened where the agent cannot alter them. An allowed action can still be a bad decision, so oversight and monitoring remain necessary.
Evidence that supports oversight
To determine what actually happened, compare the model trace with the policy decision, tool response and business record. Missing events must remain visible; a monitoring dashboard alone does not certify the control.
The Harness now brackets its six runtime steps with registration and lifecycle closure. Explore delegated authority, evaluation, human interruption, release manifests, spending limits and recourse, together with a privacy-aware telemetry design.
Open the eight-stage AI Harness Read the telemetry evidence guide
AI Harness · technical assurance reference
Component exploitation, runtime telemetry and response evidence.
Malicious instructions can enter through a retrieved document, move to an agent and trigger a connected tool. Trace the full path to find where permission should have been checked and where intervention was possible.
Explore the inter-agentic exploitation map and its Elastic SIEM / OpenTelemetry assurance field dictionary. These technical exhibits support detection engineering; they are not proof that telemetry collection or response rules are deployed in any environment.
AI DDR under Cyber Resilience
AI-DDR compares recorded behavior with approved limits. When an agent requests an unapproved tool or data scope, the alert should identify the affected process, owner and response action—not simply raise an unexplained alarm.
Inter-agentic circular threat atlas · Animated harness telemetry to Elastic · Complete integrated source reference
Source-derived evidence and clearly labeled simulation. No live enterprise connection.