agent_os_kernel 3.1.0__py3-none-any.whl

agent_control_plane/compliance.py

@@ -0,0 +1,720 @@
+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT License.
+"""
+Compliance and Regulatory Framework
+
+This module provides compliance mechanisms for various regulatory frameworks
+including EU AI Act, SOC 2, FedRAMP, GDPR, and industry-specific regulations.
+
+Research Foundations:
+    - EU AI Act (2024) - Risk-based classification and requirements
+    - SOC 2 Trust Service Criteria - Security, availability, confidentiality
+    - FedRAMP requirements for cloud service authorization
+    - GDPR Article 22 - Automated decision-making and profiling
+    - "Compliance by Design for AI Systems" (IEEE, 2024)
+    - Constitutional AI from Anthropic research
+
+See docs/RESEARCH_FOUNDATION.md for complete references.
+"""
+
+from typing import Any, Dict, List, Optional, Callable
+from dataclasses import dataclass, field
+from enum import Enum
+from datetime import datetime, timedelta
+import json
+
+
+class RegulatoryFramework(Enum):
+    """Supported regulatory frameworks"""
+    EU_AI_ACT = "eu_ai_act"
+    SOC2 = "soc2"
+    FEDRAMP = "fedramp"
+    GDPR = "gdpr"
+    HIPAA = "hipaa"
+    PCI_DSS = "pci_dss"
+    ISO27001 = "iso27001"
+
+
+class RiskCategory(Enum):
+    """EU AI Act risk categories"""
+    UNACCEPTABLE = "unacceptable"
+    HIGH_RISK = "high_risk"
+    LIMITED_RISK = "limited_risk"
+    MINIMAL_RISK = "minimal_risk"
+
+
+class ConstitutionalPrinciple(Enum):
+    """Constitutional AI principles"""
+    HELPFULNESS = "helpfulness"
+    HARMLESSNESS = "harmlessness"
+    HONESTY = "honesty"
+    TRANSPARENCY = "transparency"
+    FAIRNESS = "fairness"
+    PRIVACY = "privacy"
+
+
+@dataclass
+class ComplianceRequirement:
+    """
+    A specific compliance requirement.
+    
+    Attributes:
+        requirement_id: Unique identifier
+        framework: Which regulatory framework
+        title: Short description
+        description: Detailed requirement
+        validator: Function to check compliance
+        mandatory: Whether this is mandatory
+        control_id: Control identifier (e.g., SOC2-CC6.1)
+    """
+    requirement_id: str
+    framework: RegulatoryFramework
+    title: str
+    description: str
+    validator: Callable[[Dict[str, Any]], bool]
+    mandatory: bool = True
+    control_id: Optional[str] = None
+    metadata: Dict[str, Any] = field(default_factory=dict)
+
+
+@dataclass
+class ComplianceCheck:
+    """Result of a compliance check"""
+    compliant: bool
+    framework: RegulatoryFramework
+    checks_passed: int
+    checks_failed: int
+    failures: List[Dict[str, Any]]
+    recommendations: List[str]
+    timestamp: datetime = field(default_factory=datetime.now)
+
+
+@dataclass
+class ConstitutionalRule:
+    """
+    A constitutional rule inspired by Constitutional AI.
+    
+    These are high-level principles that guide agent behavior,
+    evaluated before and after agent actions.
+    """
+    rule_id: str
+    principle: ConstitutionalPrinciple
+    rule_text: str
+    evaluator: Callable[[str, Dict[str, Any]], float]  # Returns 0.0-1.0 compliance
+    severity: float = 1.0
+    examples: List[str] = field(default_factory=list)
+
+
+class ComplianceEngine:
+    """
+    Compliance engine for regulatory adherence.
+    
+    Features:
+    - Multi-framework compliance checking
+    - Automated control validation
+    - Audit trail generation
+    - Compliance reporting
+    - Risk assessment per EU AI Act
+    
+    Usage:
+        engine = ComplianceEngine()
+        
+        # Add requirements
+        engine.add_requirement(
+            framework=RegulatoryFramework.EU_AI_ACT,
+            title="Human oversight",
+            validator=check_human_oversight
+        )
+        
+        # Check compliance
+        result = engine.check_compliance(
+            framework=RegulatoryFramework.EU_AI_ACT,
+            context=system_context
+        )
+    """
+    
+    def __init__(self):
+        self._requirements: Dict[RegulatoryFramework, List[ComplianceRequirement]] = {}
+        self._audit_trail: List[Dict[str, Any]] = []
+        self._initialize_default_requirements()
+    
+    def add_requirement(
+        self,
+        framework: RegulatoryFramework,
+        title: str,
+        description: str,
+        validator: Callable[[Dict[str, Any]], bool],
+        mandatory: bool = True,
+        control_id: Optional[str] = None
+    ) -> str:
+        """
+        Add a compliance requirement.
+        
+        Args:
+            framework: Regulatory framework
+            title: Short description
+            description: Detailed requirement
+            validator: Function to check compliance
+            mandatory: Whether this is mandatory
+            control_id: Control identifier
+            
+        Returns:
+            requirement_id
+        """
+        import uuid
+        requirement_id = str(uuid.uuid4())
+        
+        requirement = ComplianceRequirement(
+            requirement_id=requirement_id,
+            framework=framework,
+            title=title,
+            description=description,
+            validator=validator,
+            mandatory=mandatory,
+            control_id=control_id
+        )
+        
+        if framework not in self._requirements:
+            self._requirements[framework] = []
+        
+        self._requirements[framework].append(requirement)
+        return requirement_id
+    
+    def check_compliance(
+        self,
+        framework: RegulatoryFramework,
+        context: Dict[str, Any]
+    ) -> ComplianceCheck:
+        """
+        Check compliance with a regulatory framework.
+        
+        Args:
+            framework: Framework to check against
+            context: System context for validation
+            
+        Returns:
+            ComplianceCheck with results
+        """
+        requirements = self._requirements.get(framework, [])
+        
+        passed = 0
+        failed = 0
+        failures = []
+        recommendations = []
+        
+        for req in requirements:
+            try:
+                is_compliant = req.validator(context)
+                
+                if is_compliant:
+                    passed += 1
+                else:
+                    failed += 1
+                    failures.append({
+                        "requirement_id": req.requirement_id,
+                        "title": req.title,
+                        "description": req.description,
+                        "control_id": req.control_id,
+                        "mandatory": req.mandatory
+                    })
+                    
+                    if req.mandatory:
+                        recommendations.append(
+                            f"CRITICAL: Fix mandatory requirement: {req.title}"
+                        )
+                
+                # Log to audit trail
+                self._audit_trail.append({
+                    "type": "compliance_check",
+                    "framework": framework.value,
+                    "requirement": req.title,
+                    "result": "pass" if is_compliant else "fail",
+                    "timestamp": datetime.now().isoformat()
+                })
+                
+            except Exception as e:
+                failed += 1
+                failures.append({
+                    "requirement_id": req.requirement_id,
+                    "title": req.title,
+                    "error": str(e)
+                })
+        
+        return ComplianceCheck(
+            compliant=failed == 0,
+            framework=framework,
+            checks_passed=passed,
+            checks_failed=failed,
+            failures=failures,
+            recommendations=recommendations
+        )
+    
+    def assess_risk_category(
+        self,
+        system_description: Dict[str, Any]
+    ) -> RiskCategory:
+        """
+        Assess risk category per EU AI Act.
+        
+        Args:
+            system_description: Description of the AI system
+            
+        Returns:
+            RiskCategory classification
+        """
+        # Unacceptable risk systems (Article 5)
+        unacceptable_indicators = [
+            "social_scoring",
+            "subliminal_manipulation",
+            "exploit_vulnerabilities",
+            "biometric_categorization"
+        ]
+        
+        for indicator in unacceptable_indicators:
+            if indicator in str(system_description).lower():
+                return RiskCategory.UNACCEPTABLE
+        
+        # High-risk systems (Annex III)
+        high_risk_indicators = [
+            "critical_infrastructure",
+            "education",
+            "employment",
+            "essential_services",
+            "law_enforcement",
+            "migration",
+            "justice",
+            "biometric_identification"
+        ]
+        
+        for indicator in high_risk_indicators:
+            if indicator in str(system_description).lower():
+                return RiskCategory.HIGH_RISK
+        
+        # Limited risk (transparency obligations)
+        limited_risk_indicators = [
+            "chatbot",
+            "content_generation",
+            "deepfake"
+        ]
+        
+        for indicator in limited_risk_indicators:
+            if indicator in str(system_description).lower():
+                return RiskCategory.LIMITED_RISK
+        
+        return RiskCategory.MINIMAL_RISK
+    
+    def generate_compliance_report(
+        self,
+        framework: RegulatoryFramework
+    ) -> Dict[str, Any]:
+        """
+        Generate a compliance report.
+        
+        Args:
+            framework: Framework to report on
+            
+        Returns:
+            Compliance report dictionary
+        """
+        requirements = self._requirements.get(framework, [])
+        
+        # Get recent audit trail for this framework
+        recent_checks = [
+            entry for entry in self._audit_trail
+            if entry.get("framework") == framework.value
+            and datetime.fromisoformat(entry["timestamp"]) > 
+                datetime.now() - timedelta(days=30)
+        ]
+        
+        passed = sum(1 for c in recent_checks if c["result"] == "pass")
+        failed = sum(1 for c in recent_checks if c["result"] == "fail")
+        
+        return {
+            "framework": framework.value,
+            "total_requirements": len(requirements),
+            "mandatory_requirements": sum(1 for r in requirements if r.mandatory),
+            "recent_checks": len(recent_checks),
+            "passed": passed,
+            "failed": failed,
+            "compliance_rate": (passed / len(recent_checks) * 100) if recent_checks else 0,
+            "generated_at": datetime.now().isoformat()
+        }
+    
+    def get_audit_trail(
+        self,
+        framework: Optional[RegulatoryFramework] = None,
+        days: int = 30
+    ) -> List[Dict[str, Any]]:
+        """
+        Get audit trail entries.
+        
+        Args:
+            framework: Optional framework filter
+            days: Number of days to look back
+            
+        Returns:
+            List of audit trail entries
+        """
+        cutoff = datetime.now() - timedelta(days=days)
+        
+        entries = [
+            entry for entry in self._audit_trail
+            if datetime.fromisoformat(entry["timestamp"]) > cutoff
+        ]
+        
+        if framework:
+            entries = [
+                entry for entry in entries
+                if entry.get("framework") == framework.value
+            ]
+        
+        return entries
+    
+    def _initialize_default_requirements(self):
+        """Initialize default compliance requirements"""
+        
+        # EU AI Act - High-Risk System Requirements
+        def check_human_oversight(context: Dict[str, Any]) -> bool:
+            """EU AI Act Article 14 - Human oversight"""
+            return context.get("human_oversight_enabled", False)
+        
+        self.add_requirement(
+            framework=RegulatoryFramework.EU_AI_ACT,
+            title="Human oversight",
+            description="High-risk AI systems must be designed with human oversight",
+            validator=check_human_oversight,
+            mandatory=True,
+            control_id="EU-AI-Act-Art14"
+        )
+        
+        def check_transparency(context: Dict[str, Any]) -> bool:
+            """EU AI Act Article 13 - Transparency"""
+            return context.get("provides_transparency_info", False)
+        
+        self.add_requirement(
+            framework=RegulatoryFramework.EU_AI_ACT,
+            title="Transparency and information",
+            description="Users must be informed they are interacting with an AI system",
+            validator=check_transparency,
+            mandatory=True,
+            control_id="EU-AI-Act-Art13"
+        )
+        
+        # SOC 2 - Common Criteria
+        def check_access_controls(context: Dict[str, Any]) -> bool:
+            """SOC 2 CC6.1 - Logical and physical access controls"""
+            return context.get("access_controls_implemented", False)
+        
+        self.add_requirement(
+            framework=RegulatoryFramework.SOC2,
+            title="Access controls",
+            description="Logical and physical access controls restrict access to authorized users",
+            validator=check_access_controls,
+            mandatory=True,
+            control_id="CC6.1"
+        )
+        
+        def check_monitoring(context: Dict[str, Any]) -> bool:
+            """SOC 2 CC7.2 - System monitoring"""
+            return context.get("monitoring_enabled", False)
+        
+        self.add_requirement(
+            framework=RegulatoryFramework.SOC2,
+            title="System monitoring",
+            description="The entity monitors system components and operation of controls",
+            validator=check_monitoring,
+            mandatory=True,
+            control_id="CC7.2"
+        )
+        
+        # GDPR
+        def check_data_minimization(context: Dict[str, Any]) -> bool:
+            """GDPR Article 5(1)(c) - Data minimization"""
+            collected = context.get("data_collected", [])
+            necessary = context.get("data_necessary", [])
+            return set(collected).issubset(set(necessary))
+        
+        self.add_requirement(
+            framework=RegulatoryFramework.GDPR,
+            title="Data minimization",
+            description="Personal data must be adequate, relevant and limited to what is necessary",
+            validator=check_data_minimization,
+            mandatory=True,
+            control_id="GDPR-Art5-1-c"
+        )
+
+
+class ConstitutionalAI:
+    """
+    Constitutional AI implementation for value alignment.
+    
+    Based on Anthropic's Constitutional AI approach, this provides
+    a framework for aligning agent behavior with human values through
+    explicit constitutional rules.
+    
+    Features:
+    - Define constitutional principles
+    - Evaluate actions against constitution
+    - Self-critique and revision
+    - Transparency in decision-making
+    
+    Usage:
+        constitution = ConstitutionalAI()
+        
+        # Add rules
+        constitution.add_rule(
+            principle=ConstitutionalPrinciple.HARMLESSNESS,
+            rule_text="Never assist with illegal activities",
+            evaluator=evaluate_harmlessness
+        )
+        
+        # Evaluate
+        result = constitution.evaluate("User request text", context)
+    """
+    
+    def __init__(self):
+        self._rules: List[ConstitutionalRule] = []
+        self._initialize_default_constitution()
+    
+    def add_rule(
+        self,
+        principle: ConstitutionalPrinciple,
+        rule_text: str,
+        evaluator: Callable[[str, Dict[str, Any]], float],
+        severity: float = 1.0,
+        examples: Optional[List[str]] = None
+    ) -> str:
+        """
+        Add a constitutional rule.
+        
+        Args:
+            principle: Which principle this enforces
+            rule_text: Human-readable rule description
+            evaluator: Function that evaluates compliance (0.0-1.0)
+            severity: How important this rule is
+            examples: Example applications of the rule
+            
+        Returns:
+            rule_id
+        """
+        import uuid
+        rule_id = str(uuid.uuid4())
+        
+        rule = ConstitutionalRule(
+            rule_id=rule_id,
+            principle=principle,
+            rule_text=rule_text,
+            evaluator=evaluator,
+            severity=severity,
+            examples=examples or []
+        )
+        
+        self._rules.append(rule)
+        return rule_id
+    
+    def evaluate(
+        self,
+        text: str,
+        context: Dict[str, Any]
+    ) -> Dict[str, Any]:
+        """
+        Evaluate text against constitutional rules.
+        
+        Args:
+            text: Text to evaluate (prompt, response, etc.)
+            context: Additional context
+            
+        Returns:
+            Evaluation results with compliance scores
+        """
+        results = []
+        total_score = 0.0
+        total_weight = 0.0
+        violations = []
+        
+        for rule in self._rules:
+            try:
+                compliance_score = rule.evaluator(text, context)
+                
+                results.append({
+                    "rule_id": rule.rule_id,
+                    "principle": rule.principle.value,
+                    "rule_text": rule.rule_text,
+                    "compliance_score": compliance_score,
+                    "severity": rule.severity
+                })
+                
+                total_score += compliance_score * rule.severity
+                total_weight += rule.severity
+                
+                if compliance_score < 0.7:  # Threshold for violation
+                    violations.append({
+                        "principle": rule.principle.value,
+                        "rule_text": rule.rule_text,
+                        "compliance_score": compliance_score
+                    })
+                    
+            except Exception as e:
+                results.append({
+                    "rule_id": rule.rule_id,
+                    "error": str(e)
+                })
+        
+        overall_compliance = (total_score / total_weight) if total_weight > 0 else 1.0
+        
+        return {
+            "compliant": overall_compliance >= 0.7,
+            "overall_compliance": overall_compliance,
+            "rule_results": results,
+            "violations": violations,
+            "recommendations": self._generate_recommendations(violations)
+        }
+    
+    def self_critique(
+        self,
+        proposed_response: str,
+        context: Dict[str, Any]
+    ) -> Dict[str, Any]:
+        """
+        Perform self-critique on a proposed response.
+        
+        This is a key Constitutional AI technique where the system
+        evaluates its own outputs before presenting them.
+        
+        Args:
+            proposed_response: Response to critique
+            context: Context including original prompt
+            
+        Returns:
+            Critique results with suggested revisions
+        """
+        evaluation = self.evaluate(proposed_response, context)
+        
+        revisions = []
+        if not evaluation["compliant"]:
+            for violation in evaluation["violations"]:
+                revisions.append(f"Revise to better align with: {violation['rule_text']}")
+        
+        return {
+            "approved": evaluation["compliant"],
+            "compliance_score": evaluation["overall_compliance"],
+            "violations": evaluation["violations"],
+            "suggested_revisions": revisions
+        }
+    
+    def _generate_recommendations(
+        self,
+        violations: List[Dict[str, Any]]
+    ) -> List[str]:
+        """Generate recommendations based on violations"""
+        if not violations:
+            return []
+        
+        recommendations = ["Review and revise content to address constitutional violations:"]
+        
+        for violation in violations:
+            recommendations.append(
+                f"- Align with principle: {violation['principle']} - {violation['rule_text']}"
+            )
+        
+        return recommendations
+    
+    def _initialize_default_constitution(self):
+        """Initialize default constitutional rules"""
+        
+        def evaluate_harmlessness(text: str, context: Dict[str, Any]) -> float:
+            """Evaluate harmlessness principle"""
+            harmful_keywords = [
+                "violence", "harm", "illegal", "exploit",
+                "manipulate", "deceive", "dangerous"
+            ]
+            text_lower = text.lower()
+            
+            violations = sum(1 for keyword in harmful_keywords if keyword in text_lower)
+            # Need at least 2 violations to fail
+            return max(0.0, 1.0 - (violations * 0.5))
+        
+        self.add_rule(
+            principle=ConstitutionalPrinciple.HARMLESSNESS,
+            rule_text="Avoid content that could cause harm to individuals or society",
+            evaluator=evaluate_harmlessness,
+            severity=1.0,
+            examples=[
+                "Don't provide instructions for illegal activities",
+                "Don't generate content promoting violence",
+                "Don't help with harmful manipulations"
+            ]
+        )
+        
+        def evaluate_honesty(text: str, context: Dict[str, Any]) -> float:
+            """Evaluate honesty principle"""
+            dishonest_indicators = [
+                "guarantee", "100% certain", "definitely will",
+                "impossible to", "never wrong"
+            ]
+            text_lower = text.lower()
+            
+            violations = sum(1 for indicator in dishonest_indicators if indicator in text_lower)
+            return max(0.0, 1.0 - (violations * 0.25))
+        
+        self.add_rule(
+            principle=ConstitutionalPrinciple.HONESTY,
+            rule_text="Be honest about capabilities, limitations, and uncertainty",
+            evaluator=evaluate_honesty,
+            severity=0.9,
+            examples=[
+                "Acknowledge when uncertain",
+                "Don't overstate capabilities",
+                "Be truthful about limitations"
+            ]
+        )
+        
+        def evaluate_privacy(text: str, context: Dict[str, Any]) -> float:
+            """Evaluate privacy principle"""
+            # Check for PII exposure
+            import re
+            pii_patterns = [
+                r'\b[A-Za-z0-9._%+-]+@[A-Za-z0-9.-]+\.[A-Z|a-z]{2,}\b',  # email
+                r'\b\d{3}-\d{2}-\d{4}\b',  # SSN
+                r'\b\d{3}[-.]?\d{3}[-.]?\d{4}\b',  # phone
+            ]
+            
+            violations = sum(1 for pattern in pii_patterns if re.search(pattern, text))
+            return max(0.0, 1.0 - (violations * 0.3))
+        
+        self.add_rule(
+            principle=ConstitutionalPrinciple.PRIVACY,
+            rule_text="Protect privacy and don't expose personal information",
+            evaluator=evaluate_privacy,
+            severity=1.0,
+            examples=[
+                "Don't include PII in responses",
+                "Respect data confidentiality",
+                "Follow privacy best practices"
+            ]
+        )
+    
+    def get_constitution(self) -> List[Dict[str, Any]]:
+        """Get the full constitution as a readable format"""
+        return [
+            {
+                "principle": rule.principle.value,
+                "rule_text": rule.rule_text,
+                "severity": rule.severity,
+                "examples": rule.examples
+            }
+            for rule in self._rules
+        ]
+
+
+def create_compliance_suite() -> Dict[str, Any]:
+    """
+    Create a complete compliance suite with multiple frameworks.
+    
+    Returns:
+        Dictionary with compliance engine and constitutional AI
+    """
+    return {
+        "compliance_engine": ComplianceEngine(),
+        "constitutional_ai": ConstitutionalAI()
+    }