AI Control
Govern employee AI use with Moftgate AI Control: prompt capture, allow, redact and block policies, data classifications and shadow AI discovery.
What is AI Control#
Employees use chat assistants, coding assistants and AI-enabled desktop apps every day. When a customer list or source code is pasted into a prompt, the data leaves the organization. AI Control lets you see this use and control it.
The Moftgate agent inspects prompts sent to monitored AI providers on the device. Based on your policy it sends the prompt, masks sensitive data, blocks it, or denies the application any access to the provider. Every decision is recorded. The module works the same in SaaS and On-Prem deployments.
The AI Control entry in the left menu has these pages:
| Page | What it shows |
|---|---|
| Overview | KPI strip, day × hour heatmap, enforcement funnel, user → application → provider flow and risk ranking. |
| Prompt Activity | Every captured prompt: user, device, application, provider, model, action and matched classification. |
| Policies | Application × provider rules and their actions. |
| Applications | IDEs, CLIs, browsers and desktop apps, with process names per operating system. |
| Providers | Web assistants and API backends with their domains. Approved providers are marked here. |
| Data Classifications | Patterns, keywords and validators that rules match in prompts and uploads. |
| Shadow AI | Use of unapproved or uncatalogued providers. |
How prompt capture works#
The agent inspects traffic to AI providers through the Web Content Gateway proxy. It parses the prompt per provider, compares it with the rules and applies the decision. Records stay between the agent and your own server.
| Mode | When | Behavior |
|---|---|---|
| Full mode | A DLP license exists and the gateway is on in the profile. | Proxy, SSL decryption and Inline Web DLP must be on. |
| AI-only mode | No full mode, and at least one enabled AI Control policy applies to the device. | Only AI provider domains are decrypted; other traffic passes untouched. It is the only capture path in SaaS. |
| Off | No enabled policy and no full mode. | No prompts are recorded. |
Creating the first policy#
- 1Open Policies
Go to AI Control > Policies and click New policy.
- 2Choose the scope
Select the devices the policy applies to. Capture starts on those devices.
- 3Add rules
For each rule choose an application or application kind, a provider, optional classifications or keywords, and an action.
- 4Save and order
Enable the policy. Drag rows to change priority if needed. Agents receive the change right away.
Policies and actions#
A rule defines an application × provider match and an action. An empty application or provider list means “all”. You can choose an application kind (browser, IDE, CLI, desktop app) instead of single apps; apps of that kind added to the catalog later are covered automatically.
| Action | Result |
|---|---|
| Allow & log | The prompt is sent and recorded. |
| Redact | Matched data is masked and the masked prompt is sent. Needs at least one classification or keyword. If a match remains after masking, the request is blocked. |
| Block | The prompt is not sent and the event is recorded. |
| Deny app | The application may not talk to the provider at all. |
- Policies are evaluated in priority order; the first policy whose rule matches decides.
- A rule with classifications or keywords applies only when the prompt contains one of them. A rule without them applies to all prompts of the selected apps and providers.
- Keywords are plain text, not regex; case and Turkish i/ı/İ/I differences are ignored. A rule takes up to 100 keywords.
- A rule that points to a disabled or deleted catalog item is dropped; it never widens to “all”.
Application and provider catalog#
The Providers page lists the web assistants and API backends the agent inspects, with their domains. The Applications page attributes AI traffic to an IDE, CLI, browser or desktop app by process name. The catalog ships with real domains and process names.
Data classifications#
The Data Classifications page holds the patterns rules match in prompts and uploads: regex, keywords and validators (for example Luhn-checked card numbers, Turkish ID and tax numbers). Built-in templates cover common data types. The list is shared with DLP, so you define a classification once.
Shadow AI#
Every prompt sent to an unapproved or uncatalogued provider appears on the Shadow AI page. To approve the use, turn on the provider’s Sanctioned switch on the Providers page or use the Sanction action on the Shadow AI page.
For risks and recommended policies per provider, see the Shadow AI guide.
Overview and Prompt Activity#
Overview summarizes AI use for the selected date range. Without a date it shows the last 30 days. Provider and application filters apply to every chart.
- KPI strip: prompts, flagged rate, unsanctioned share, active users and sensitive-data matches, each compared with the previous period.
- Enforcement funnel: captured → policy matched → redacted → blocked → denied app.
- Risk ranking and agentic AI data access: which app sent which classification to which provider.
- Clicking a chart opens Prompt Activity with the same filter.
Permissions and privacy#
| Permission | What it allows |
|---|---|
| ai_governance.view | View pages and lists. |
| ai_governance.manage | Manage policies, providers and applications; see prompt content and full user names. |
Prompts are stored up to 4,000 characters and stay on your own server. When a policy changes, the affected agents receive the new rules right away.