Skip to main content

Concepts

Concepts

Explanations of how AIControls works. Read these to understand the model before changing configuration — for step-by-step instructions, see Tasks.

Managed GitHub App connection

Consent, repository scope, readiness, and recovery for the staged managed connection.

The LLM providers AIControls sends requests to — Anthropic, OpenAI, Azure OpenAI, Google Vertex, Amazon Bedrock, and OpenAI-compatible vendors.

How a request reaches a provider: model patterns, Access Binding grants that name a route, credential mode, strategies, failover, and fail-closed errors.

Identity & Activity

Identity types, risk scores, behavioral baselines, sessions, session replay, adoption analytics, and the audit log.

Cost Governance

Budgets, rules, limits, alerts, and context optimization — including compression and lossless retrieval.

The Home Cost dashboard: 24-hour, 7-day, and 30-day windows, which totals follow the window, the reporting time zone, and partial results.

Security & Compliance

Security posture, detections, content safety scanning, skill risk scoring, compliance coverage across five frameworks including the EU AI Act and NIST AI RMF, and the AI Analyst's generated narratives explaining what's driving a score.

What the scanner detects — credentials, PII, and prompt-injection/jailbreak attempts — and how the optional ML classifier fits in.

Runtime Controls

Kernel-level observation of AI workloads in your own Kubernetes cluster — what reached a provider without going through governance, and what the evidence does and does not prove.

Workload Protection

Register → Observe → Enforce, with independent Registration, Runtime, Enforcement, and Verification statuses.

Workload Blueprints and Registrations

How a Workload Registration, an identity binding, and a Workload Blueprint combine into the live access and runtime policy for a workload.

Workload access

Take a workload from identity binding to Audit, recommendations, Promote, Enforce, and containment, and find the workloads whose calls no access policy covers.

Policy anatomy, enforcement modes, matching rules, and the built-in policy library.

The CSA NIST T1–T4 autonomy model, the default tier, and how the built-in policies scale enforcement to it.

Approval Workflow

The request → approve → grant → expire/revoke lifecycle, and single- vs multi-approver co-sign.

How MCP tool-call proxying, logging, and authorization work.

Map IdP teams to MCP servers with Access Bindings, enforced at token issuance and per tool call.

Named, reusable tool sets across MCP servers — AIControls' virtual MCP servers, compiled onto the same policy plane.

Network Egress Filtering

How raw outbound HTTP(S) calls from agent processes are intercepted, governed, and audited.

What developers see in their own portal — budgets, scores, tokens, and sessions.

AI delivery attribution — linking merged pull requests back to the AI sessions that produced them.

Effectiveness Leaderboard

Complexity-adjusted AI effectiveness scores with configurable score bands and evidence behind each result.

Configuration Model

Deployment settings (Helm/config.yaml) vs. app settings (Settings UI) — seeding, upgrades, and drift notifications.

Plans and Entitlements

How a signed license sets your plan, add-ons, and limits, and where the Free tier and the deprecated config-file tier fit.