AI Governance Intelligence
for the Agentic Era

Who This Guide Is For

Risk Tiers

Every AI system in KoraSafe is classified into one of four risk tiers under the EU AI Act. The tier determines the controls, documentation, and monitoring required.

Autonomy Levels

Every registered AI system is assigned an autonomy level that defines the governance controls it requires and how Guardian Agents enforce boundaries.

The Six Pillars of AI Governance

KoraSafe measures governance maturity across six pillars. Each pillar has four to five controls that map to regulatory requirements and industry standards.

Regulatory Frameworks

Running an Assessment

Navigate to Risk Assessment and select a registered AI system (or start a new one from the landing page). The assessment works through five sections:

Step 1: Agent type

Select all functions that apply to your AI system. Most agents serve multiple functions, and each one can trigger different regulatory requirements. For example, a hiring-screening agent that also uses biometric recognition triggers both Employment (Annex III, 4a) and Biometric Categorization (Annex III, 1a) high-risk categories under the EU AI Act simultaneously.

Step 2: Industry

Select the industry your agent serves, not the industry your company operates in. A fraud-detection agent for a bank operates in Financial Services even if the vendor builds fintech software.

Step 3: Data categories

Select every data type the system accesses, processes, or generates. Health data triggers GDPR special categories (Art. 9). Biometric data can trigger EU AI Act prohibitions. Select all that apply; the engine resolves overlapping requirements automatically.

Step 4: Affected Populations

Include everyone your system makes decisions about, interacts with, or communicates to. Vulnerable groups (minors, patients, people with disabilities) trigger stricter requirements under both EU AI Act and GDPR. Select all groups that may be reached, not just the primary intended users.

Step 5: Jurisdictions and Existing Controls

Select where your system operates or serves users. Then work through the governance posture section: for each practice (e.g., "Human review before consequential decisions," "Bias testing in production"), select your current status. Honest answers produce better recommendations.

Assessment Output

Follow-Up Q&A

After the assessment completes, a chat interface appears at the bottom of the results page. Ask KoraSafe follow-up questions grounded in your specific assessment context, for example, "What does Art. 9(2)(g) mean for my health-data system?" or "Which controls address the DPIA requirement?" Answers are cited against the regulatory knowledge base.

Adding a System

Click Add System in the AI Registry. The minimum required fields are:

• Name, a clear, internal identifier (e.g., "Customer Churn Predictor v2")
• System type, classification, prediction, generative, agentic, or hybrid
• Domain, the business area or industry the system serves
• Underlying model, the model family or vendor (e.g., GPT-4o, Llama 3, internal model)
• Owner, the team member accountable for this system's governance
• Autonomy level, Observe, Advise, Supervised, or Autonomous
• Lifecycle status, Development, Staging, Production, Deprecated

Fleet View

The fleet view shows all registered systems in a sortable, filterable table. Filter by risk classification, autonomy level, lifecycle status, or owner. Click any system to open its detail view.

System Detail Tabs

Governance Heatmap

The heatmap displays a matrix of all registered AI systems (rows) against the governance dimensions (columns). Color coding shows status at a glance:

Maturity Radar

The maturity radar scores your organization across seven governance dimensions against a five-level scale:

Agent Evals (EDD Pipeline)

The Eval-Driven Development (EDD) pipeline applies a quality gate to AI systems before they reach production. Four stages:

Evals are scored across six weighted dimensions: Accuracy, Bias, Hallucination Rate, Safety, Compliance, and Latency. A composite score must exceed the configured threshold before the Gate stage passes.

Policy Types

KoraSafe supports seven enforcement policy types. Policies apply to individual systems or system groups:

Violation Management

When a guardian agent or policy rule detects a violation, it appears in the Violations tab with severity, timestamp, affected system, and the specific rule triggered. Filter by severity or status. Resolve violations through the admin workflow, each resolution is timestamped and logged to the immutable audit trail.

Guardian Controls Panel

Navigate to Enforcement > Guardian Controls to activate, pause, or configure individual Guardian Agents. Each guardian shows its trigger count, last-active timestamp, and current autonomy level. You can adjust each guardian's autonomy level independently, a guardian set to Observe reports findings without taking action, while one set to Act can execute remediation automatically.

How Guardians Work

Each guardian operates as a specialist agent that evaluates inputs, outputs, or system behavior against its specialized detection rules. Results are returned as structured findings with severity, evidence, and recommended remediation. At autonomy level Observe, findings are logged only. At Propose, the guardian creates a pending resolution that a human approves. At Act, the guardian executes remediation directly and logs the action.

PII Types Detected

Knowledge Base

KoraSafe's answers are grounded in KoraSafe's regulatory knowledge base, a continuously updated corpus covering:

Citing Sources

Every KoraSafe response includes source citations, the specific document, article, or section the answer draws from. Citations appear as footnotes in the answer. Click a citation to view the full source text. If KoraSafe cannot find regulatory support for a claim, it says so rather than speculating.

Assessment Follow-Up Mode

After completing a risk assessment, the KoraSafe interface enters assessment context mode. Questions asked here are answered with awareness of your specific system's profile, risk tier, and governance posture. For example, asking "What do I need to fix first?" yields a prioritized, system-specific remediation list rather than generic guidance.

Compliance Checklist

The checklist organizes 24 core governance controls across the Six Pillars. Each control has a status: Done, In Progress, or Not Started. Progress bars at the top of each pillar show overall completion.

Controls are mapped to regulatory requirements, so checking one off automatically updates the compliance posture displayed in the Governance heatmap and assessment results.

RACI Matrix

The RACI matrix defines accountability for every governance control. The four roles:

Edit any cell to assign a team member to a role. Changes are saved automatically and logged to the audit trail. Export the completed RACI as PDF for regulatory submissions or board reporting.

Exporting for Audits

Both the checklist and RACI matrix export to CSV and PDF. The PDF export includes:

• Organization name and export timestamp
• Overall completion percentage per pillar
• Control status with last-updated dates
• Assigned responsible and accountable owners

Connected Services

Document Ingestion

Import regulatory documents into the knowledge base from Integrations > Document Ingestion. Provide the document title, full text or URL, category (law, guidance, enforcement action, framework), and jurisdiction. KoraSafe AI-indexes the document and makes it available to KoraSafe for citations within a few minutes.

Setting Up Slack

Roles and Permissions

Inviting Users

Navigate to Admin > Users > Invite. Enter the email address and select a role. KoraSafe sends a token-based invitation link valid for 48 hours. The invitee clicks the link, completes account setup, and is added to your organization automatically.

SSO and MFA

KoraSafe supports SAML 2.0 and OIDC for Single Sign-On. Configure your identity provider (Okta, Azure AD, Google Workspace) in Admin > Security > SSO. Provide your IdP metadata URL or upload the XML metadata file. Once SSO is configured, you can enforce it organization-wide, new users must authenticate through your IdP.

Multi-factor authentication can be enforced for all users or specific roles from Admin > Security > MFA. TOTP (authenticator app) and SMS are supported.

API Key Management

Create scoped API keys from Admin > API Keys. Each key can be scoped to specific endpoints (e.g., read-only access to assessments, or write access to the AI Registry). Rotate or revoke keys at any time. All key usage is logged with timestamps, IP addresses, and endpoint calls to the audit trail.

Feature Flags

Owners can enable or disable platform features per organization from Admin > Feature Flags. This allows gradual rollout of new modules to your team, or disabling features not yet ready for your governance workflow.

Agent Card

KoraSafe publishes an A2A (Agent-to-Agent) agent card at /.well-known/agent.json. Other AI agents use this card to discover KoraSafe's capabilities and establish governed connections.

Available MCP Endpoints

Authentication

All MCP API calls require a bearer token. Create a scoped API key in Admin > API Keys and include it in the Authorization header:

CI/CD Integration Example

To gate a deployment on governance status, add a step to your pipeline that calls /api/assess and fails the build if the compliance score is below your threshold:

MCP Dashboard

Monitor MCP server activity, request volume, error rates, and latency from the MCP Dashboard at /dashboard. Use this to verify integrations are working and debug failed requests.

Running a code audit

There are four ways to run a code audit:

• Web platform — Navigate to Insights and Operations > Code Audit. Drag-drop a source file or paste a GitHub URL. The Code Auditor agent scans the code and maps findings to regulatory controls.
• KoraSafe agent bar — Type /code-audit followed by a description (e.g., /code-audit the hiring agent). Findings render as structured cards with Apply fix, Reject, and Escalate actions.
• GitHub Action — Add korasafe/kora-action to your workflow. The action runs on every pull request, posts findings as PR comments, and creates a Check Run that blocks merges when critical violations are found.
• VS Code extension — Diagnostics appear on file save. Open the KoraSafe sidebar for a compliance score ring and severity-grouped findings.

Audit findings dashboard

Navigate to Insights and Operations > Audit Findings to see all findings across surfaces in a unified view. Filter by severity (critical, high, medium, low), category (PII, HITL, secrets, data handling), status (open, acknowledged, resolved), and source (web, CI/CD, IDE, browser).

Select multiple findings and use bulk status update to acknowledge or resolve them in a single action.

Slash commands for audit

• /code-audit — Run the Code Auditor agent on an agent or code snippet
• /deps — Run the Dependency Auditor to scan packages for CVEs and license issues
• /fix — Generate remediation patches for a specific finding

Browsing the catalog

Navigate to Insights and Operations > Policy Packs. The catalog shows available packs organized by regulation (EU AI Act, GDPR, HIPAA, US state laws). Each pack displays its current version, last updated date, and subscriber count.

Subscribing and pinning

Click Subscribe on a pack to begin enforcing its policies across your surfaces. By default, subscriptions auto-update when a new pack version is published. To lock to a specific version, toggle Pin version — this ensures your policies remain stable during critical periods.

Human review gate

When a regulation is amended and a pack version is bumped, the update is held for human review before enforcement begins. An admin must approve the new version in the review queue. This prevents untested regulatory changes from reaching production.

Surface distribution

Each policy can be toggled per surface: web platform, CI/CD, IDE, and browser. This lets you enforce a policy in CI/CD (blocking deployments) while leaving it in advisory mode in the IDE (showing diagnostics without blocking saves).

VS Code extension

Install the .vsix package from Extensions > Install from VSIX in VS Code. Once installed:

• Diagnostics on save — Squiggly underlines appear on governance violations (PII, missing HITL gates, hardcoded secrets) every time you save a file.
• Sidebar — The KoraSafe sidebar (shield icon) shows a compliance score ring and findings grouped by severity.
• Quick fixes — Click the lightbulb on a diagnostic to apply a one-click fix.
• Hover tooltips — Hover over a finding to see which regulation and article it maps to.
• Commands — KoraSafe: Scan File, KoraSafe: Scan Workspace, KoraSafe: Set API Key from the command palette.

Chrome browser extension

Load the extension unpacked from extensions/chrome/ in chrome://extensions. The extension monitors AI chat interfaces and LLM API calls:

• Network interception — Detects API calls to OpenAI, Anthropic, Azure OpenAI, Google AI, Cohere, and Bedrock endpoints.
• Shadow AI detection — Identifies when team members use ChatGPT, Gemini, Copilot, or other AI tools on unauthorized pages.
• PII scanning — Scans text entered into AI chat inputs for SSNs, email addresses, phone numbers, and other PII before it leaves the browser.
• Side panel — Click the extension icon to open the side panel with Summary, Findings, and Timeline tabs.

GitHub Action

Add the KoraSafe governance action to any repository:

The action collects changed files in the PR, submits them to the Code Auditor and Dependency Auditor agents, and posts a findings summary as a PR comment. Set fail-on to critical, high, or medium to control the merge-blocking threshold.

GitLab CI

Include the reusable template in your .gitlab-ci.yml:

JS and Python SDKs

For programmatic access, use the official SDKs:

Cost center management

Navigate to Insights and Operations > FinOps. Create cost centers by team, project, or use case. Allocate monthly budgets and track actual spend against each center in real time.

Budget alerts

Set threshold-based alerts (e.g., "alert when 80% of budget consumed"). Alerts route through the severity-based notification system — critical budget breaches trigger Slack and email notifications.

Cost-per-action tracking

Break down LLM spend by action type: KoraSafe queries, risk assessments, guardian scans, code audits, and document ingestion. Identify which workflows consume the most tokens and optimize accordingly.

Reports

• Usage forecast — Project future spend based on current trends to plan capacity.
• Chargeback report — Allocate AI costs to business units for internal billing.
• Value report — Quantify governance ROI: compliance gaps closed, findings remediated, audit hours saved.

Navigate to Insights and Operations > System Health. The dashboard shows:

• Service probes — Every 5 minutes, health probes check the database, auth provider, LLM endpoints, and integrations. Results display as green (operational), amber (degraded), or red (down).
• Error log — Searchable log with structured error codes, request IDs, and timestamps. Stack traces and internal paths are never exposed to API clients.
• Endpoint health — Per-endpoint latency and error rate metrics. Spot degraded routes before they impact users.
• Database health — Connection pool utilization, query performance, and table size monitoring.
• SLA report — Track availability, response times, and governance finding resolution windows against your SLA targets.

How alert routing works

When a governance finding is detected, KoraSafe routes the alert based on severity:

Critical alert banner

When critical or high-severity alerts are active, a persistent banner appears at the top of the platform. The banner polls every 60 seconds and links directly to the audit findings dashboard. Banners can be dismissed per session but reappear until the underlying alert is resolved.

Notification preferences

Navigate to your profile or Administration > Notification Preferences. Toggle delivery channels (in-app, email, Slack) independently for each category:

• Governance — Risk assessments, compliance gaps, policy violations, governance score changes
• Agents — Pending approvals, trust tier changes, circuit breakers, error spikes
• Security — Failed logins, new device access, role changes, API key expiry
• Usage — Approaching limits, budget thresholds, overages, plan renewals
• System — Service degradation, error rate spikes, SLA breaches
• Team — New members, invitations accepted, role changes

Preferences are saved per user and per organization.

Alert rules engine

Admins can define custom alert rules with a metric, operator, threshold, cooldown period, and channel routing. Rules are evaluated every 5 minutes by the cron engine. Stale alerts (open for more than 24 hours with no recurrence) are auto-resolved.

SLA compliance

Every alert tracks first_detected_at, acknowledged_at, and resolved_at. The SLA compliance endpoint (/api/alerts/sla) returns breach rates per severity tier, helping teams identify bottlenecks in their remediation workflows.