Policy Configuration

Define fine-grained access control policies for AI agents with condition-based rules, escalation paths, and compliance enforcement.

Overview

Agent policies enable organizations to enforce governance rules at the agent level, controlling what actions agents can perform and under what conditions.

Source: ow-ai-backend/models_agent_registry.py

Compliance: NIST AC-3, PCI-DSS 7.1, HIPAA 164.312(a)

Policy Structure

AgentPolicy Model

Field	Type	Description
policy_name	String	Human-readable name
policy_description	Text	Purpose description
is_active	Boolean	Enabled status
priority	Integer	Evaluation order (lower = higher)
conditions	JSONB	When policy applies
policy_action	String	What happens
action_params	JSONB	Action configuration

Policy Actions

Action	Description	Use Case
allow	Permit the action	Low-risk operations
block	Deny the action	Prohibited operations
require_approval	Queue for human review	Medium-risk operations
escalate	Send to security team	High-risk operations

Creating Policies

Example: Block High-Risk Transactions

{
  "policy_name": "Block High-Risk Transactions",
  "policy_description": "Prevent autonomous agents from processing high-value transactions",
  "is_active": true,
  "priority": 10,
  "conditions": {
    "action_type": "transaction",
    "risk_above": 70,
    "agent_type": "autonomous"
  },
  "policy_action": "block",
  "action_params": {
    "notification": "security-team@company.com",
    "audit_level": "critical"
  }
}

Example: Require Approval for PII Access

{
  "policy_name": "PII Access Approval",
  "policy_description": "Require human approval for any PII data access",
  "is_active": true,
  "priority": 20,
  "conditions": {
    "resource_type": "pii",
    "action_type": ["read", "write", "delete"]
  },
  "policy_action": "require_approval",
  "action_params": {
    "approvers": ["compliance-team"],
    "timeout_hours": 24,
    "escalate_on_timeout": true
  }
}

Example: Escalate Database Deletes

{
  "policy_name": "Database Delete Escalation",
  "policy_description": "Escalate all database delete operations",
  "is_active": true,
  "priority": 5,
  "conditions": {
    "action_type": "database_delete"
  },
  "policy_action": "escalate",
  "action_params": {
    "escalate_to": "security-team",
    "notification_channel": "slack",
    "webhook_url": "https://hooks.slack.com/..."
  }
}

Condition Syntax

Available Condition Fields

Field	Type	Description
action_type	string/array	Action type filter
agent_type	string	Agent classification
risk_above	integer	Minimum risk score
risk_below	integer	Maximum risk score
resource_type	string	Target resource type
time_window	object	Time-based restrictions
data_classification	string/array	Data sensitivity

Condition Operators

Operator	Description	Example
Direct value	Exact match	"action_type": "transaction"
Array	Any match	"action_type": ["read", "write"]
Range	Numeric comparison	"risk_above": 70
Object	Complex condition	"time_window": {"outside": "09:00-17:00"}

Policy Priority

Policies evaluated in priority order (lower number = higher priority):

Priority 1:  Block Critical Actions
Priority 10: Require Dual Approval
Priority 20: Standard Approval
Priority 100: Default Allow

Priority Guidelines

Priority Range	Use Case
1-10	Critical security blocks
11-50	High-priority restrictions
51-100	Standard governance
101-500	Organizational rules
501+	Default/fallback policies

Action Parameters

Block Action Params

{
  "notification": "security@company.com",
  "audit_level": "critical",
  "log_reason": true,
  "alert_on_repeated": true
}

Require Approval Params

{
  "approvers": ["team-leads", "security"],
  "timeout_hours": 24,
  "escalate_on_timeout": true,
  "require_justification": true,
  "min_approvers": 1
}

Escalate Params

{
  "escalate_to": "security-team",
  "notification_channel": "slack",
  "webhook_url": "https://...",
  "include_context": true,
  "severity": "high"
}

Agent-Level Policies

Per-Agent Configuration

Each registered agent can have specific policies:

{
  "agent_id": "financial-advisor-001",
  "policies": [
    {
      "policy_name": "Transaction Limit",
      "conditions": {
        "action_type": "transaction",
        "amount_above": 10000
      },
      "policy_action": "require_approval"
    }
  ]
}

Agent Policy Inheritance

1. Global organization policies (highest priority)
2. Agent-type policies
3. Agent-specific policies
4. Default allow (lowest priority)

Compliance Mapping

NIST AC-3 (Access Enforcement)

Control	Implementation
AC-3(1)	Role-based policy assignment
AC-3(2)	Dual authorization support
AC-3(3)	Mandatory access control

PCI-DSS 7.1 (Restrict Access)

Requirement	Implementation
7.1.1	Policy-based access restrictions
7.1.2	Least privilege enforcement
7.1.3	Role-based permissions

API Reference

Endpoint	Method	Description
/api/governance/policies	GET	List all policies
/api/governance/policies	POST	Create policy
/api/governance/policies/{id}	PUT	Update policy
/api/governance/policies/{id}	DELETE	Delete policy
/api/governance/policies/evaluate	POST	Test policy

Best Practices

1. Start restrictive: Begin with block/require_approval, relax as needed
2. Use meaningful names: Policy names should explain purpose
3. Document conditions: Add descriptions for complex conditions
4. Test before deploy: Use evaluate endpoint to test policies
5. Regular review: Audit policies quarterly for relevance
6. Version policies: Use policy_description for change notes

Troubleshooting

Policy not triggering

Check:

• Is policy is_active: true?
• Do conditions match the action?
• Is priority correct (not overridden)?

Actions being blocked unexpectedly

Solution: Check higher-priority policies; use evaluate endpoint.

Multiple policies conflicting

Solution: Review priority order; first matching policy wins.

---

Source: models_agent_registry.py