Guard checks text against your policies, masks GCC personal data and writes each decision to an audit trail, before the text reaches a model.
Your AI policy is a document. Nothing checks a prompt against it.
National IDs, phones and IBANs travel with the question.
The auditor asks what was checked. Nobody can show it.
Sends the prompt or the answer to Guard.
Applies your policies, the content filter and the personal-data scan. Writes an audit row.
Allow, warn or block, with the masked text and the reason.
Field names from the Guard API. Anar supplies no models. Guard checks the text that your application sends to it.
Policy
Type: keyword. Action: block.
Keywords: Project Falcon, مشروع الصقر
Rules are keyword lists or patterns. Each rule logs, warns or blocks. The strictest result wins.
Output of Guard's policy check on these prompts, run 18 November 2026.
A metric, a threshold and a time window. For example, more than 20 blocks in 15 minutes.
Guard raises it once, then waits for the cooldown. Your team acknowledges it.
Your SIEM or chat tool gets the alert, signed with HMAC-SHA256, with retries.
| Time | Event | Severity | User | Result |
|---|---|---|---|---|
| 09:14:02 | policy_violation | critical | finance.analyst | block: Classified project names |
| 09:14:40 | pii_detection | warning | hr.officer | 3 entities masked |
| 09:15:11 | evaluation | info | citizen.portal | allowed |
Example rows. The columns are the ones Guard exports: id, timestamp, event type, content snippet, result, user, model and severity.
Filter by event type, severity, user and time.
Download the trail as CSV or JSON for your auditor.
A chained record that shows any later edit.
An event for each conversation turn.
An event for each transcription.
An event for each run.
The fields each product sends to Guard. The values are an example. Set the Guard address in each product to turn it on.
| Capability | Status | Note |
|---|---|---|
| Policies, personal-data masking, audit export | Shipped | Keyword and pattern rules. GCC IDs, phones, IBANs, cards, passports. |
| Alerts and signed webhooks | Shipped | Thresholds, cooldowns and retries. |
| Safety classifier | Partial | Uses a model you configure. It marks text safe when that model does not answer. |
| Inline proxy to Anar Gateway | Partial | Built. No Anar product routes through it yet. |
| Framework reports | Partial | Five frameworks. The score counts activity. It does not test controls. |
| Names in Arabic text, tenants, tamper evidence | Roadmap | One customer per deployment today. |
- Keyword and pattern policies
- GCC personal-data masking
- Audit trail with export
- Alerts and signed webhooks
- Safety classifier inside your boundary
- Arabizi and Gulf dialect detection
- Names and places in Arabic text
- Human review queue
- Policy versions and replay
- Evidence mapped to each control
A one-month pilot: we deploy Guard in your environment and connect it to one team's AI use.
anarlabs.ai/contact