AI Governance Policies (Pro)

An AI Governance Policy says whether an AI component is allowed: authorized, unauthorized, or needs review. A policy can apply to every Asset, to one Organization, or to one Asset. DefectDojo evaluates every component in the AI Inventory against the policies that apply to it. An unauthorized component becomes an ordinary Finding, so priority, SLAs, the Triage Engine, tickets, notifications, dashboards and reports all apply to it.

AI Governance Policies are part of the AI Inventory and Governance feature (beta). Turn it on from the Feature Flags page.

Creating a Policy#

Go to Settings > Finding Workflow > AI Governance Policies and choose New Policy. A policy has:

A global policy needs a superuser. An Organization or Asset policy needs edit rights on that Organization or Asset. Every change is recorded in the audit log.

DefectDojo ships four starter policies, all disabled. They allow the MCP SDKs, allow the major AI provider SDKs, flag every MCP server until reviewed, and flag model files checked into a repository. Enable the ones that fit, or use them as examples.

How a Decision Is Made#

When several policies match one component on one Asset:

1. The most specific scope wins. An Asset policy beats an Organization policy, which beats a global one.
2. Within that scope, the most specific match wins. An exact Package URL beats a pattern, and a longer pattern beats a shorter one. A pattern beats a provider, and a provider beats a category.
3. A remaining tie goes to the stricter decision. Unauthorized beats needs review, which beats authorized.

A component that no enabled policy matches needs review. DefectDojo never treats unknown AI as allowed.

Decisions are re-evaluated in the background after every inventory import and every policy change, for exactly the Assets the change can affect. A nightly pass catches anything else. A superuser can also choose Re-evaluate all on the policies page.

Findings#

An unauthorized component raises one Finding:

• It is titled Unauthorized AI component: name.
• It is tagged ai-governance, plus ai-<kind> (for example ai-mcp-server).
• It lives in an AI Governance engagement and test on the Asset.
• Its description names the policy, its justification, and where in the repository the component was found.

Re-evaluation keeps it in step with the repository:

• Authorizing the component, or removing it from the repository, mitigates the Finding on the next evaluation.
• Unauthorizing it again reopens the same Finding. It never creates a second one.
• A Finding someone marked false positive, out of scope or risk accepted is left alone.

Components that need review raise nothing by default and are counted on the inventory page. To triage them like any other Finding, turn on Raise Info Findings for AI components needing review at the top of the policies page. Each undecided component then raises an Info Finding, and turning the setting off mitigates them.

Automating the Response#

Because violations are Findings, the Triage Engine handles them with no new setup. A template ships in the Triage Engine: AI Governance: ticket and notify on an unauthorized AI component. It triggers on new Findings tagged ai-governance at High or Critical, opens a ticket and sends an alert. Adopt it, point the ticket at your project, and try it in Simulate first.

A notification event, AI component discovered, fires the first time a kind of AI component appears on an Asset, such as its first MCP server. Removing and re-adding a component does not fire it again. Choose where it goes in your notification settings.

Using the API#

Policies and decisions are available with an API token:

GET  /api/v2/ai_governance/policies/
POST /api/v2/ai_governance/policies/
GET  /api/v2/ai_governance/verdicts/?product=12&decision=unauthorized
POST /api/v2/ai_governance/policies/evaluate/   {"product": 12}