Aller au contenu principal

Policy Configuration

Policies are the core enforcement mechanism in Noxys. They define which AI interactions are allowed, coached, or blocked based on real-time conditions.

Policy Fundamentals​

A policy consists of:

ComponentPurposeExample
NameHuman-readable description"Block PII on ChatGPT"
DescriptionOptional context"Prevents sensitive data sharing on US-based services per policy #20260315"
ConditionsWhen the rule applies (AND logic)platform_id == "chatgpt" AND risk_score > 0.8
ActionWhat happens on matchBlock / Coach / Log / Redact
PriorityExecution order (lower = first)10
EnabledActive or disabled toggleOn / Off

Creating Policies​

Step 1: Access Policy Creation​

  1. Go to Policies in the sidebar
  2. Click + New Policy
  3. Fill out the form

Step 2: Basic Information​

FieldRequiredNotes
NameYesUnique, max 200 chars. Example: "Block PII on ChatGPT"
DescriptionNoMax 500 chars. Include compliance reference if applicable
EnabledNoDefaults to off. Toggle on to enforce immediately
PriorityNo0-1000 (default: 100). Lower = evaluated first

Step 3: Add Conditions​

Conditions determine when a policy applies. All conditions must match (AND logic).

To add a condition:

  1. Click + Add Condition
  2. Select a field from dropdown
  3. Choose an operator
  4. Enter a value
  5. Click Add

Repeat to add multiple conditions (all must be true for policy to trigger).

Step 4: Choose Action​

Select one of four actions:

ActionBehaviorWhen to Use
BlockInteraction is preventedFor high-risk or non-compliant scenarios
CoachWarning appears, user can send anywayFor educational nudges
LogSilently logged, no UI shownFor monitoring and audit trails
RedactRisk Signals are replaced with [REDACTED:TYPE] tokens before the prompt is sent — no user confirmationFor frictionless compliance: let users keep working while removing sensitive data transparently

Redact action details:

  • Runs before submission, in the browser extension, with zero network round-trip.
  • Every detected Risk Signals span is replaced in place with a placeholder: alice@acme.fr → [REDACTED:EMAIL], FR76 3000 6000 0112 3456 7890 189 → [REDACTED:IBAN], etc.
  • Supported subtypes today: EMAIL, PHONE, IBAN, CREDIT_CARD, NIR_FR, SIRET_FR, VAT_EU, IP_ADDRESS, API_KEY, PASSWORD.
  • The redacted prompt is what the AI service receives. The original is never stored or logged (raw text never leaves the browser).
  • Overlap rules apply: a digit span that is both a potential phone number and an IBAN is redacted once, as the IBAN (more specific match wins).
  • Idempotent: running redact on already-redacted text is a no-op.

Step 5: Review & Create​

  1. Review all details
  2. Click Create Policy

The policy is created but disabled by default (good for testing).

  1. Toggle Enabled to activate

Policy Conditions Reference​

Available Fields​

Platform Identifier​

Field: platform_id Operators: eq (equals), neq (not equals), in (list), nin (not in list) Values: String (platform name)

Examples:

platform_id eq "chatgpt"
platform_id neq "claude"
platform_id in ["chatgpt", "gemini"]
platform_id nin ["deepseek", "grok"]

Supported platforms: chatgpt, claude, gemini, deepseek, perplexity, mistral, copilot, poe, huggingchat, grok, mammouth, typingmind, openrouter, windsurf, cursor

Risk Score​

Field: risk_score Operators: gt (>), gte (≥), lt (<), lte (≤), eq (=) Values: Float (0-1)

Interpretation:

  • 0.0 = No Risk Signals detected
  • 0.1-0.3 = Low risk
  • 0.3-0.7 = Medium risk
  • 0.7-0.9 = High risk
  • 0.9-1.0 = Critical risk

Examples:

risk_score gte 0.8          # Block high-risk only
risk_score gt 0.5 # Block medium+ risk
risk_score eq 0 # Block only if no Risk Signals (inverse logic)
risk_score lt 0.2 # Block only if low risk (unusual)

Direction​

Field: direction Operators: eq Values: "outbound" (user sending to AI) or "inbound" (AI response)

Examples:

direction eq "outbound"     # Monitor user inputs
direction eq "inbound" # Monitor AI responses (rare use)

Note: Currently, only outbound (prompts) are monitored. Inbound support coming in v0.2.

Interaction Type​

Field: interaction_type Operators: eq Values: "prompt", "completion", "tool_call", "embedding"

Examples:

interaction_type eq "prompt"      # Block prompts only
interaction_type eq "tool_call" # Block function calls only

User ID / Email​

Field: user_id or user_email Operators: eq, neq, contains Values: UUID, email address, or wildcard pattern

Examples:

user_email eq "alice@acme.fr"
user_email contains "@finance.acme.fr"
user_id eq "uuid-123-456"

Department / Group​

Field: department (requires SSO) Operators: eq, neq, in, nin Values: String (from Entra ID, LDAP, SAML)

Examples:

department eq "Finance"
department in ["Finance", "Legal", "HR"]
department neq "Engineering"

Requires: SSO enabled (Entra ID, LDAP, SAML, OIDC)

Classification Count​

Field: classification_count Operators: gte (≥), gt (>), lt (<), lte (≤), eq Values: Integer (0+)

Examples:

classification_count gte 1      # Any Risk Signal detected
classification_count gte 2 # Multiple Risk Signals types
classification_count gt 3 # More than 3 Risk Signal matches

Classification Types​

Field: classification_types Operators: contains, not_contains, intersects, not_intersects Values: String (Risk Signals type name) or list

Examples:

classification_types contains "EMAIL"                    # Contains email
classification_types contains "CREDIT_CARD" # Contains credit card
classification_types intersects ["EMAIL", "PHONE"] # Contains any of these
classification_types not_contains "FR_NIR" # Excludes social security

Available types: EMAIL, PHONE, CREDIT_CARD, IBAN, FR_NIR, FR_SIRET, FR_SIREN, MEDICAL_TERM, LEGAL_REFERENCE, API_KEY, IP_ADDRESS

Source​

Field: source Operators: eq, neq, in Values: "browser_extension", "proxy", "endpoint_agent", "api"

Examples:

source eq "browser_extension"
source in ["browser_extension", "api"]

Data Region​

Field: data_region (requires service mapping) Operators: eq, neq, in Values: "EU", "US", "APAC", "UNKNOWN"

Examples:

data_region neq "EU"        # Non-EU services
data_region eq "US" # US-based services only

Example Policies​

1. Block All DeepSeek Usage​

Use case: Organization policy prohibits Chinese AI services.

Policy Name: Block DeepSeek
Description: Organizational compliance - prohibited vendor
Enabled: Yes
Priority: 10
Action: Block

Conditions:
- platform_id eq "deepseek"

Effect: Any attempt to use DeepSeek is blocked. User sees red banner.


2. Coach Users on Risk Signals (Non-Blocking)​

Use case: Educate users about data sharing without blocking work.

Policy Name: Coach on Sensitive Data
Description: Nudge users to review Risk Signals before sending
Enabled: Yes
Priority: 20
Action: Coach

Conditions:
- classification_count gte 1
- risk_score gte 0.5

Effect: Any message with Risk Signals gets a yellow warning. User can still send.


3. Block Risk Signals on US-Based Services​

Use case: Comply with EU data sovereignty (no US cloud).

Policy Name: Block Risk Signals on US Services
Description: GDPR compliance - prevent US exposure
Enabled: Yes
Priority: 15
Action: Block

Conditions:
- data_region eq "US"
- classification_types contains "EMAIL"
OR
- data_region eq "US"
- classification_types contains "PHONE"
OR
- data_region eq "US"
- classification_types contains "CREDIT_CARD"

Effect: Emails, phone numbers, and credit cards are blocked on ChatGPT, Gemini, etc.


4. Block Finance Team from Non-EU Services​

Use case: Department-specific policy (requires SSO).

Policy Name: Finance - EU Services Only
Description: Finance dept restricted to EU cloud providers
Enabled: Yes
Priority: 5
Action: Block

Conditions:
- department eq "Finance"
- data_region neq "EU"

Effect: Finance team members cannot use non-EU AI services.


5. Log All API Key Usage​

Use case: Detect potential credential leaks.

Policy Name: Log API Keys
Description: Monitor API key leakage attempts
Enabled: Yes
Priority: 30
Action: Log

Conditions:
- classification_types contains "API_KEY"

Effect: Silent logging. Admins can review in Interactions later.


6. Block High-Risk Medical Data (Healthcare)​

Use case: HIPAA compliance.

Policy Name: Block Medical Data on Unsecured Services
Description: Protect patient PII per HIPAA
Enabled: Yes
Priority: 5
Action: Block

Conditions:
- classification_types contains "MEDICAL_TERM"
- source eq "browser_extension"

Effect: Patient names, diagnoses, etc., are blocked from public AI services.


7. Restrict Specific Users to Approved Services​

Use case: Contractor or partner access control.

Policy Name: Contractor - Claude Only
Description: External partner limited to Claude
Enabled: Yes
Priority: 10
Action: Block

Conditions:
- user_email contains "@partner-corp.com"
- platform_id neq "claude"

Effect: Partner users can only access Claude. All other services blocked.


8. Redact Risk Signals Transparently (Zero-Friction Compliance)​

Use case: Let employees keep using public AI services for productivity, but strip all Risk Signals before it leaves the browser. No coaching overlay, no user decision, no friction.

Policy Name: Auto-Redact Risk Signals on Public AI
Description: Transparent redaction — user keeps working, data stays safe
Enabled: Yes
Priority: 25
Action: Redact

Conditions:
- classification_count gte 1
- data_region neq "EU"

Effect: A user typing "Send follow-up to alice@acme.fr about IBAN FR76 3000 6000 0112 3456 7890 189" into ChatGPT will see the prompt silently rewritten to "Send follow-up to [REDACTED:EMAIL] about IBAN [REDACTED:IBAN]" before submission. The assistant's response is still useful ("I'll draft a follow-up note…") but no Risk Signals leaves the machine.

When to prefer Redact over Coach:

  • High-volume teams where pop-ups would interrupt work every few minutes.
  • Compliance-first orgs that want guaranteed enforcement without relying on user choice.
  • Situations where the user is expected to keep raw data in their own system of record, not the AI.

When NOT to use Redact:

  • When the prompt only makes sense with the real data (legal drafting, medical records — use Block).
  • When you want users to learn what Risk Signals look like (use Coach first, Redact later).

9. French Social Security Numbers - Strict Block​

Use case: French GDPR compliance (FR_NIR is sensitive).

Policy Name: Block FR Social Security (NIR)
Description: FR_NIR never allowed per CNIL guidance
Enabled: Yes
Priority: 1
Action: Block

Conditions:
- classification_types contains "FR_NIR"

Effect: Any mention of French social security number blocks the interaction.

Policy Evaluation Order​

When an AIInteraction arrives at the backend:

1. Load all ENABLED policies
2. Sort by PRIORITY (ascending: 0, 5, 10, 15, ...)
3. For each policy:
a. Evaluate all conditions with AND logic
b. If all conditions match:
- Apply action (Block/Coach/Log)
- Stop evaluation (don't check remaining policies)
- Create audit record
c. If any condition fails:
- Move to next policy
4. If no policies matched:
- Log with action = "Allowed"

Example Walkthrough​

Scenario: User alice@acme.fr sends a message with an email address to ChatGPT.

Interaction Data:

platform_id: "chatgpt"
user_email: "alice@acme.fr"
department: "Finance"
risk_score: 0.25
classification_types: ["EMAIL"]
classification_count: 1
data_region: "US"

Policies (enabled, by priority):

  1. Priority 5: "Finance - EU Services Only" (Block if Finance + non-EU)
  2. Priority 10: "Block Risk Signals on US Services" (Block if US + EMAIL/PHONE)
  3. Priority 20: "Coach on Sensitive Data" (Coach if Risk Signals + risk > 0.5)

Evaluation:

StepPolicyConditionResultAction
1Finance - EUdepartment eq "Finance" ✓ AND data_region neq "EU" ✓MATCHBlock
——(Stop here, don't check Policy 2 or 3)——

Result: Message is Blocked by "Finance - EU Services Only".

Admin sees: Policy violation, policy ID, timestamp.


Editing Policies​

Modify a Policy​

  1. Go to Policies
  2. Click the policy name
  3. Click Edit
  4. Change any field:
    • Name, description
    • Add/remove conditions
    • Change action
    • Adjust priority
  5. Click Save

Effect: Updated policy takes effect immediately for all future interactions.

Historical interactions are unchanged (their decision record is immutable).

Enable/Disable Without Editing​

  1. Go to Policies
  2. Find the policy in the list
  3. Click the Enabled toggle

Off = policy not evaluated. On = policy evaluated.

Useful for:

  • Testing new policies
  • Temporarily pausing a rule during investigation
  • A/B testing effectiveness

Duplicate a Policy​

  1. Click a policy
  2. Click ⋮ More → Duplicate
  3. Edit the copy (it's disabled by default)
  4. Click Save

Useful for creating similar policies with slight variations.

Deleting Policies​

  1. Click the policy
  2. Click Delete
  3. Confirm: "Are you sure? This action is permanent."

Important: Deleting a policy does NOT delete historical records. All past interactions remain in the audit log forever.

After deletion: New interactions skip the deleted policy during evaluation.

Policy Best Practices​

1. Start Restrictive, Relax Over Time​

Bad: Create vague policies that block half your usage.

Good: Start with Log-only policies to understand usage patterns. Graduate to Coach. Finally, Block only clear violations.

Week 1: Log all DeepSeek usage
→ Understand how much is used
Week 2: Coach on Risk Signals (non-blocking)
→ Train users
Week 3: Block PII on ChatGPT
→ Enforce compliance

2. Use Descriptive Names​

Bad: "Policy 1", "Rule A"

Good: "Block PII on ChatGPT", "Coach Finance on US Services"

Names should tell you the intent without reading conditions.

3. Document Context in Description​

Name: Block DeepSeek
Description: Organizational decision per security review 2026-03-15.
Complies with EU AI Act Article 4 (prohibited practices).
Approved by: Security Committee (Chair: Jane Doe)

4. Manage Priority Carefully​

Lower-priority (lower number) policies should be most important:

Priority 1:  Block FR_NIR (GDPR critical)
Priority 5: Block Risk Signals on US (data sovereignty)
Priority 10: Block DeepSeek (organizational policy)
Priority 20: Coach on High Risk (educational)
Priority 100: Log all (default)

5. Test with Log-Only First​

Never immediately Block. Always test:

  1. Create policy with Action = Log
  2. Run for 1-2 days
  3. Check Interactions to see what would be blocked
  4. If acceptable, change Action to Block
  5. Enable

6. Avoid Overly Complex Conditions​

Bad: 8 conditions, hard to understand logic

Good: 2-3 clear conditions, one policy per intent

If you need complex logic, split into multiple policies with different priorities.

7. Use Negation Carefully​

Confusing:

NOT platform_id = "claude"

Clear:

platform_id in ["chatgpt", "gemini"]

Positive conditions are easier to understand.

Common Policy Patterns​

Pattern 1: Restrict by Vendor​

Block specific vendors (e.g., Chinese, Russian):
platform_id in ["deepseek", "grok"]
Action: Block

Pattern 2: Protect Sensitive Departments​

Block non-EU services for Finance/Legal:
department in ["Finance", "Legal"]
data_region neq "EU"
Action: Block

Pattern 3: Education Over Enforcement​

Coach on any Risk Signal:
classification_count gte 1
Action: Coach

Pattern 4: Compliance-Driven​

Block specific Risk Signals types:
classification_types contains "FR_NIR"
Action: Block

Pattern 5: High-Risk Only​

Block only high-risk interactions:
risk_score gte 0.8
Action: Block

Monitoring Policy Effectiveness​

Check Trigger Count​

In the Policies list, see "Triggered" column (last 7 days):

  • 0 triggers = Policy never matched (too strict, or rare scenario)
  • 1-10 triggers = Healthy enforcement
  • 100+ triggers = May need adjustment (too broad?)

Review False Positives​

If users complain policies are too strict:

  1. Go to Interactions
  2. Filter by policy decision: "Blocked"
  3. Review a sample
  4. Identify patterns (e.g., "always blocks on certain keyword")
  5. Adjust condition (e.g., increase risk_score threshold)

A/B Test Policies​

  1. Create new policy with Coach (instead of Block)
  2. Run for 1 week
  3. Compare behavior change
  4. Decide to upgrade to Block or revert

Next Steps​


Need help?

  • Email: support@noxys.eu
  • Template policies: Available in console under Policy Templates
  • Audit trail: All policy changes logged in Audit Log