Yeetek's picture
Upload 43 files
7d5083d verified
"""
Input validation and safety checks for prompts.
This module provides comprehensive validation utilities for prompt safety,
content filtering, and input sanitization to ensure secure and reliable
prompt processing.
"""
import re
import html
from typing import List, Dict, Optional, Tuple, Set
from dataclasses import dataclass
from enum import Enum
from config.logging import get_logger
logger = get_logger(__name__)
class ValidationSeverity(str, Enum):
"""Severity levels for validation issues."""
INFO = "info"
WARNING = "warning"
ERROR = "error"
CRITICAL = "critical"
@dataclass
class ValidationIssue:
"""Represents a validation issue."""
severity: ValidationSeverity
code: str
message: str
location: Optional[str] = None
suggestion: Optional[str] = None
@dataclass
class ValidationResult:
"""Result of validation process."""
is_valid: bool
issues: List[ValidationIssue]
sanitized_content: Optional[str] = None
risk_score: float = 0.0
@property
def has_errors(self) -> bool:
"""Check if there are any error-level issues."""
return any(issue.severity in [ValidationSeverity.ERROR, ValidationSeverity.CRITICAL]
for issue in self.issues)
@property
def has_warnings(self) -> bool:
"""Check if there are any warning-level issues."""
return any(issue.severity == ValidationSeverity.WARNING for issue in self.issues)
class PromptValidator:
"""
Comprehensive prompt validation and safety checking.
Provides multiple layers of validation including:
- Content safety and injection detection
- Format compliance checking
- Length and structure validation
- Business context validation
"""
# Prompt injection patterns
INJECTION_PATTERNS = {
"role_manipulation": [
r"(?i)\b(ignore|forget|disregard)\s+(previous|above|earlier|all)\s+(instructions?|prompts?|rules?)",
r"(?i)\b(act\s+as|pretend\s+to\s+be|roleplay\s+as|simulate)\s+(?!a\s+business|an?\s+analyst)",
r"(?i)\b(you\s+are\s+now|from\s+now\s+on|instead)\s+",
r"(?i)\b(override|bypass|disable|turn\s+off)\s+"
],
"system_commands": [
r"(?i)\b(system|admin|root|sudo)\s*[\(\[]",
r"(?i)\b(exec|eval|run|execute)\s*[\(\[]",
r"(?i)<script|javascript:|data:|vbscript:",
r"(?i)\$\{|\$\(|`[^`]*`" # Variable expansion
],
"format_breaking": [
r"(?i)\b(don't\s+use|ignore|skip|avoid)\s+(json|format|structure)",
r"(?i)\b(plain\s+text|free\s+form|unstructured|raw\s+output)",
r"(?i)\b(no\s+format|without\s+format|format\s+free)"
],
"data_extraction": [
r"(?i)\b(show|reveal|display|print|output)\s+(all|your|the)\s+(data|information|content)",
r"(?i)\b(what\s+is\s+your|tell\s+me\s+your)\s+(system|prompt|instructions?)",
r"(?i)\b(dump|export|leak)\s+"
]
}
# Suspicious keywords by category
SUSPICIOUS_KEYWORDS = {
"high_risk": {
"jailbreak", "prompt_injection", "ignore_instructions", "system_override",
"admin_access", "root_privileges", "bypass_safety", "disable_filters"
},
"medium_risk": {
"eval", "exec", "sudo", "admin", "root", "shell", "command",
"injection", "exploit", "hack", "bypass"
},
"format_risk": {
"plain_text", "no_json", "unstructured", "free_form", "raw_output",
"ignore_format", "skip_structure", "format_free"
}
}
# Required elements for topic extraction prompts
REQUIRED_ELEMENTS = {
"output_format": ["json", "format", "structure"],
"topic_fields": ["topic_name", "topic_type", "confidence_score"],
"business_context": ["business", "topic", "extract", "analyze"]
}
# Content length limits
LENGTH_LIMITS = {
"min_prompt_length": 20,
"max_prompt_length": 15000,
"max_line_length": 500,
"max_word_length": 50
}
def __init__(self):
"""Initialize the prompt validator."""
self.logger = get_logger(f"{__name__}.{self.__class__.__name__}")
# Compile regex patterns for performance
self._compiled_patterns = {}
for category, patterns in self.INJECTION_PATTERNS.items():
self._compiled_patterns[category] = [re.compile(pattern) for pattern in patterns]
def validate_prompt(self, prompt: str, context: Optional[Dict] = None) -> ValidationResult:
"""
Comprehensive prompt validation.
Args:
prompt: The prompt text to validate
context: Optional context for validation
Returns:
ValidationResult with issues and sanitized content
"""
issues = []
risk_score = 0.0
try:
# Basic validation
basic_issues, basic_risk = self._validate_basic_structure(prompt)
issues.extend(basic_issues)
risk_score += basic_risk
# Safety validation
safety_issues, safety_risk = self._validate_safety(prompt)
issues.extend(safety_issues)
risk_score += safety_risk
# Format validation
format_issues, format_risk = self._validate_format_compliance(prompt)
issues.extend(format_issues)
risk_score += format_risk
# Business context validation
business_issues, business_risk = self._validate_business_context(prompt, context)
issues.extend(business_issues)
risk_score += business_risk
# Content sanitization
sanitized_content = self._sanitize_content(prompt)
# Determine overall validity
is_valid = not any(issue.severity in [ValidationSeverity.ERROR, ValidationSeverity.CRITICAL]
for issue in issues)
# Normalize risk score (0.0 to 1.0)
risk_score = min(1.0, max(0.0, risk_score))
result = ValidationResult(
is_valid=is_valid,
issues=issues,
sanitized_content=sanitized_content,
risk_score=risk_score
)
self.logger.debug(
f"Prompt validation completed: valid={is_valid}, "
f"issues={len(issues)}, risk_score={risk_score:.2f}"
)
return result
except Exception as e:
self.logger.error(f"Error during prompt validation: {str(e)}")
return ValidationResult(
is_valid=False,
issues=[ValidationIssue(
severity=ValidationSeverity.CRITICAL,
code="VALIDATION_ERROR",
message=f"Validation failed: {str(e)}"
)],
risk_score=1.0
)
def _validate_basic_structure(self, prompt: str) -> Tuple[List[ValidationIssue], float]:
"""Validate basic prompt structure and length."""
issues = []
risk_score = 0.0
# Length validation
if len(prompt) < self.LENGTH_LIMITS["min_prompt_length"]:
issues.append(ValidationIssue(
severity=ValidationSeverity.ERROR,
code="PROMPT_TOO_SHORT",
message=f"Prompt is too short ({len(prompt)} chars). Minimum: {self.LENGTH_LIMITS['min_prompt_length']}",
suggestion="Provide more detailed instructions for better results"
))
risk_score += 0.3
if len(prompt) > self.LENGTH_LIMITS["max_prompt_length"]:
issues.append(ValidationIssue(
severity=ValidationSeverity.WARNING,
code="PROMPT_TOO_LONG",
message=f"Prompt is very long ({len(prompt)} chars). May exceed token limits.",
suggestion="Consider breaking into smaller, focused prompts"
))
risk_score += 0.1
# Line length validation
lines = prompt.split('\n')
long_lines = [i for i, line in enumerate(lines)
if len(line) > self.LENGTH_LIMITS["max_line_length"]]
if long_lines:
issues.append(ValidationIssue(
severity=ValidationSeverity.INFO,
code="LONG_LINES",
message=f"Found {len(long_lines)} very long lines",
location=f"Lines: {long_lines[:5]}", # Show first 5
suggestion="Consider breaking long lines for better readability"
))
# Word length validation
words = prompt.split()
long_words = [word for word in words
if len(word) > self.LENGTH_LIMITS["max_word_length"]]
if len(long_words) > 5: # Allow some long words
issues.append(ValidationIssue(
severity=ValidationSeverity.INFO,
code="LONG_WORDS",
message=f"Found {len(long_words)} very long words",
suggestion="Very long words may indicate encoded content or errors"
))
# Character encoding validation
try:
prompt.encode('utf-8')
except UnicodeEncodeError as e:
issues.append(ValidationIssue(
severity=ValidationSeverity.ERROR,
code="ENCODING_ERROR",
message=f"Invalid character encoding: {str(e)}",
suggestion="Ensure prompt uses valid UTF-8 characters"
))
risk_score += 0.2
return issues, risk_score
def _validate_safety(self, prompt: str) -> Tuple[List[ValidationIssue], float]:
"""Validate prompt safety and detect injection attempts."""
issues = []
risk_score = 0.0
# Check for injection patterns
for category, patterns in self._compiled_patterns.items():
matches = []
for pattern in patterns:
found = pattern.findall(prompt)
matches.extend(found)
if matches:
severity = ValidationSeverity.CRITICAL if category == "system_commands" else ValidationSeverity.WARNING
risk_increase = 0.4 if category == "system_commands" else 0.2
issues.append(ValidationIssue(
severity=severity,
code=f"INJECTION_{category.upper()}",
message=f"Detected potential {category.replace('_', ' ')} injection",
location=f"Matches: {matches[:3]}", # Show first 3 matches
suggestion="Review prompt for unintended injection patterns"
))
risk_score += risk_increase
# Check for suspicious keywords
prompt_lower = prompt.lower()
for risk_level, keywords in self.SUSPICIOUS_KEYWORDS.items():
found_keywords = [kw for kw in keywords if kw in prompt_lower]
if found_keywords:
if risk_level == "high_risk":
severity = ValidationSeverity.ERROR
risk_increase = 0.3
elif risk_level == "medium_risk":
severity = ValidationSeverity.WARNING
risk_increase = 0.1
else: # format_risk
severity = ValidationSeverity.WARNING
risk_increase = 0.15
issues.append(ValidationIssue(
severity=severity,
code=f"SUSPICIOUS_{risk_level.upper()}",
message=f"Found {len(found_keywords)} {risk_level.replace('_', ' ')} keywords",
location=f"Keywords: {found_keywords[:5]}",
suggestion="Review keywords for potential security risks"
))
risk_score += risk_increase
# Check for HTML/XML content
if '<' in prompt and '>' in prompt:
html_tags = re.findall(r'<[^>]+>', prompt)
if html_tags:
issues.append(ValidationIssue(
severity=ValidationSeverity.WARNING,
code="HTML_CONTENT",
message=f"Found {len(html_tags)} HTML-like tags",
location=f"Tags: {html_tags[:3]}",
suggestion="HTML content may indicate injection attempts"
))
risk_score += 0.1
# Check for excessive special characters
special_chars = re.findall(r'[^\w\s\.\,\!\?\;\:\-\(\)\[\]\{\}\"\'\/\\]', prompt)
if len(special_chars) > len(prompt) * 0.1: # More than 10% special chars
issues.append(ValidationIssue(
severity=ValidationSeverity.INFO,
code="EXCESSIVE_SPECIAL_CHARS",
message=f"High ratio of special characters ({len(special_chars)}/{len(prompt)})",
suggestion="Excessive special characters may indicate encoded content"
))
return issues, risk_score
def _validate_format_compliance(self, prompt: str) -> Tuple[List[ValidationIssue], float]:
"""Validate that prompt enforces proper output format."""
issues = []
risk_score = 0.0
prompt_lower = prompt.lower()
# Check for required format elements
missing_categories = []
for category, required_words in self.REQUIRED_ELEMENTS.items():
found_words = [word for word in required_words if word in prompt_lower]
if not found_words:
missing_categories.append(category)
if missing_categories:
issues.append(ValidationIssue(
severity=ValidationSeverity.WARNING,
code="MISSING_FORMAT_ELEMENTS",
message=f"Missing format elements: {missing_categories}",
suggestion="Ensure prompt requests proper JSON format and required fields"
))
risk_score += 0.1 * len(missing_categories)
# Check for format-breaking instructions
format_breaking_patterns = [
r"(?i)\b(don't|do\s+not|avoid|skip)\s+(use\s+)?json",
r"(?i)\b(plain\s+text|free\s+form|unstructured)",
r"(?i)\b(ignore|skip|avoid)\s+(format|structure)"
]
for pattern in format_breaking_patterns:
if re.search(pattern, prompt):
issues.append(ValidationIssue(
severity=ValidationSeverity.ERROR,
code="FORMAT_BREAKING_INSTRUCTION",
message="Prompt contains instructions that may break output format",
suggestion="Remove instructions that discourage structured output"
))
risk_score += 0.3
break
# Check for JSON format enforcement
json_indicators = ["json", "format", "structure", "array", "object"]
found_indicators = sum(1 for indicator in json_indicators if indicator in prompt_lower)
if found_indicators < 2:
issues.append(ValidationIssue(
severity=ValidationSeverity.INFO,
code="WEAK_FORMAT_ENFORCEMENT",
message="Prompt may not strongly enforce JSON output format",
suggestion="Add explicit JSON format requirements"
))
return issues, risk_score
def _validate_business_context(
self,
prompt: str,
context: Optional[Dict] = None
) -> Tuple[List[ValidationIssue], float]:
"""Validate business context and topic extraction focus."""
issues = []
risk_score = 0.0
prompt_lower = prompt.lower()
# Check for business-relevant keywords
business_keywords = [
"business", "topic", "extract", "analyze", "insight", "category",
"customer", "client", "requirement", "feedback", "solution"
]
found_business_keywords = sum(1 for kw in business_keywords if kw in prompt_lower)
if found_business_keywords < 3:
issues.append(ValidationIssue(
severity=ValidationSeverity.INFO,
code="LIMITED_BUSINESS_CONTEXT",
message="Prompt may lack sufficient business context",
suggestion="Include more business-focused instructions for better results"
))
# Check for topic extraction focus
topic_keywords = ["topic", "theme", "subject", "category", "segment"]
found_topic_keywords = sum(1 for kw in topic_keywords if kw in prompt_lower)
if found_topic_keywords == 0:
issues.append(ValidationIssue(
severity=ValidationSeverity.WARNING,
code="NO_TOPIC_FOCUS",
message="Prompt doesn't explicitly mention topic extraction",
suggestion="Add explicit topic extraction instructions"
))
risk_score += 0.1
# Validate context if provided
if context:
if context.get("language") and context["language"] not in prompt_lower:
issues.append(ValidationIssue(
severity=ValidationSeverity.INFO,
code="LANGUAGE_MISMATCH",
message="Prompt language may not match specified context language",
suggestion="Ensure prompt language matches context requirements"
))
return issues, risk_score
def _sanitize_content(self, prompt: str) -> str:
"""Sanitize prompt content while preserving functionality."""
try:
# HTML escape potentially dangerous characters
sanitized = html.escape(prompt, quote=False)
# Remove null bytes and control characters (except newlines and tabs)
sanitized = re.sub(r'[\x00-\x08\x0B\x0C\x0E-\x1F\x7F-\x9F]', '', sanitized)
# Normalize whitespace
sanitized = re.sub(r'\s+', ' ', sanitized)
sanitized = sanitized.strip()
return sanitized
except Exception as e:
self.logger.error(f"Error sanitizing content: {str(e)}")
return prompt # Return original if sanitization fails
def validate_template_variables(self, variables: Dict[str, str]) -> ValidationResult:
"""Validate template variables for safety."""
issues = []
risk_score = 0.0
for var_name, var_value in variables.items():
# Validate variable name
if not re.match(r'^[a-zA-Z_][a-zA-Z0-9_]*$', var_name):
issues.append(ValidationIssue(
severity=ValidationSeverity.WARNING,
code="INVALID_VARIABLE_NAME",
message=f"Variable name '{var_name}' contains invalid characters",
suggestion="Use only alphanumeric characters and underscores"
))
# Validate variable value
if isinstance(var_value, str):
var_validation = self.validate_prompt(var_value)
if var_validation.risk_score > 0.5:
issues.append(ValidationIssue(
severity=ValidationSeverity.WARNING,
code="RISKY_VARIABLE_VALUE",
message=f"Variable '{var_name}' contains potentially risky content",
suggestion="Review variable content for safety"
))
risk_score += 0.1
return ValidationResult(
is_valid=not any(issue.severity == ValidationSeverity.ERROR for issue in issues),
issues=issues,
risk_score=min(1.0, risk_score)
)
def get_safety_recommendations(self, validation_result: ValidationResult) -> List[str]:
"""Get safety recommendations based on validation results."""
recommendations = []
if validation_result.risk_score > 0.7:
recommendations.append("Consider rewriting the prompt to reduce security risks")
if validation_result.has_errors:
recommendations.append("Fix all error-level issues before using the prompt")
if validation_result.has_warnings:
recommendations.append("Review and address warning-level issues")
# Specific recommendations based on issue codes
issue_codes = {issue.code for issue in validation_result.issues}
if "INJECTION_ROLE_MANIPULATION" in issue_codes:
recommendations.append("Remove instructions that attempt to change the AI's role")
if "FORMAT_BREAKING_INSTRUCTION" in issue_codes:
recommendations.append("Ensure prompt enforces structured JSON output")
if "PROMPT_TOO_SHORT" in issue_codes:
recommendations.append("Provide more detailed instructions for better results")
if "NO_TOPIC_FOCUS" in issue_codes:
recommendations.append("Add explicit topic extraction instructions")
# Add general security recommendation for high-risk prompts
if any("INJECTION" in code or "SUSPICIOUS" in code for code in issue_codes):
recommendations.append("Review prompt for potential security risks")
return recommendations
# Global validator instance
_validator: Optional[PromptValidator] = None
def get_prompt_validator() -> PromptValidator:
"""Get or create global prompt validator instance."""
global _validator
if _validator is None:
_validator = PromptValidator()
return _validator
def validate_prompt_safety(prompt: str, context: Optional[Dict] = None) -> ValidationResult:
"""Convenience function for prompt safety validation."""
validator = get_prompt_validator()
return validator.validate_prompt(prompt, context)
def sanitize_prompt_content(prompt: str) -> str:
"""Convenience function for prompt content sanitization."""
validator = get_prompt_validator()
return validator._sanitize_content(prompt)