Policies and Detection
Configure monitor and active policies, decisions, approvals, detector patterns, and safe policy testing.
Plan requirement: Monitor evaluates built-in detection in Monitor mode. Protect adds active policy enforcement, templates, and approvals. Business and Enterprise add custom content analysis and file inspection.
Policies determine how Orthalon responds when sensitive content or unapproved AI applications are detected across your organization's endpoints and model traffic.
1. Policy Decisions
| Decision | Action Taken |
|---|---|
| Allow | Continues normally with metadata logged for audit and compliance. |
| Coach / Warn | Displays an inline warning modal explaining company policy and risk before user confirmation. |
| Redact | Automatically replaces sensitive values with synthetic tokens before transmission. |
| Require Approval | Pauses the request and routes metadata to administrators for async authorization. |
| Block | Immediately stops the prompt, upload, or API request and logs the violation. |
- Monitor Mode: Records what the policy decision would have been without altering user actions.
- Active Mode: Enforces live decisions (Block, Redact, Coach, Require Approval) in real time.
2. Creating and Ordering Policies
- Navigate to Controls → Policies.
- Start from an enterprise template or create a custom policy.
- Choose Monitor mode for initial evaluation, then switch to Active mode once validated.
- Set matching conditions: application domain, sanction state (
SANCTIONED,UNSANCTIONED,UNDER_REVIEW), detector class, or employee identity context. - Select the desired decision action.
- Rules are evaluated top-down by priority. High-priority, specific rules must be placed above general fallback rules.
3. Identity-Aware Conditions
| Condition | Required Configuration | Scope |
|---|---|---|
| Application | Normalized domain (e.g. chatgpt.com, claude.ai) | Supported browser sites, desktop proxies, and gateway routes |
| Sanction State | Application classification in Discovery | SANCTIONED, UNSANCTIONED, UNDER_REVIEW |
| Account Type | Managed profile or adapter account signal | Distinguishes corporate from personal AI accounts |
| Corporate Domain | Managed domain configuration (e.g. company.com) | Matches authorized enterprise SSO domains |
| Detector Class | Built-in or custom sensitive data detectors | Scans prompt content and file attachments |
| Organization Role | Member role assignment in Orthalon | Targets specific teams, groups, or administrative tiers |
4. Built-in Detectors and Content Analysis
Business and Enterprise administrators can configure matchers and tune false-positive exclusions under Controls → Content Analysis:
| Detector | Accepted Signal | Validation & False-Positive Control |
|---|---|---|
| US Social Security Number | 9-digit valid SSN format | Rejects invalid area and serial ranges |
| UK National Insurance (NINO) | 2 letters, 6 digits, 1 letter beside label | Requires prefix keyword (NI, NINO) |
| Brazilian CPF | 11 digits beside CPF label | Validates official modulo-11 check digits |
| Polish PESEL | 11 digits beside PESEL label | Validates official checksum and calendar date |
| NHS Number | 10 digits beside NHS label | Validates official modulus-11 check digit |
| API & Cloud Keys | AWS, GCP, OpenAI, Anthropic, GitHub, Stripe | High-entropy regex and provider format verification |
| Source Code & Repos | Proprietary code markers, functions, secrets | Language syntax and pattern heuristics |
| Internal IPv4 / CIDR | Custom company subnets (e.g. 10.0.0.0/8) | Bounded IP range matching |
5. Approval Workflow
When a rule triggers Require Approval:
- The request is securely paused on the local endpoint or gateway.
- A metadata-only record (application, detector types, timestamp, requester) appears under Controls → Approval Requests.
- Administrators approve or deny the request in real time via the dashboard.
- Upon approval, the local sensor atomically releases the exact held payload without storing raw prompt text on the server.
6. Optional Redacted Prompt Evidence
Enterprise organizations can retain the post-redaction text that users actually send. This setting is disabled by default and is intended for investigations, policy tuning, and employee-risk review without collecting the original detected values.
- Open Operations → Prompt Evidence as an organization owner, administrator, or security administrator.
- Enable collection and choose the environments where it is permitted: browser, desktop, AI gateway, and explicitly wrapped MCP stdio tools.
- Set a retention period from 7 to 90 days, then save. Updated policy is delivered automatically to enrolled sensors and services.
- Open an individual record to reveal its retained text. Every reveal is recorded in the organization audit log.
Orthalon stores only a bounded copy of content released by a Redact decision. Detected values have already been replaced with typed placeholders such as <EMAIL_1> or <AWS_KEY_1>. Original detected values, blocked requests, allowed requests, coached requests, approval payloads, and file bodies are not retained by this feature. The evidence list exposes metadata first; authorized reviewers must deliberately reveal one record at a time.
Examples of evidence an authorized reviewer may see:
Email the revised renewal proposal to <EMAIL_1> and copy <EMAIL_2> after legal approves it.Use <AWS_KEY_1> for the staging deployment, then notify <EMAIL_1> when the smoke test completes.The customer reported SSN <US_SSN_1> and card <CREDIT_CARD_1> in the support transcript. Remove both before summarizing the case.Review this authentication flow. The embedded credential was replaced with <PRIVATE_KEY_1> before the request left the protected environment.
Placeholders identify the detector type and occurrence. They do not contain the original value and cannot be used to reconstruct it.
Disabling collection stops new records. Existing records remain encrypted until their retention period ends unless an authorized administrator permanently purges them from the Prompt Evidence page.