Keyword Detection

HipoCap’s Prompt Guard (Stage 1) uses specialized models to detect suspicious patterns and keywords in function calls and results. This provides fast, low-latency detection before more expensive LLM analysis.

What is Keyword Detection?

Keyword detection identifies sensitive patterns and keywords in function inputs and outputs. It’s part of Stage 1 (Input Analysis) and provides fast protection against sensitive data exposure. Common keywords detected:

Security keywords (confidential, classified, top secret)
Business keywords (proprietary, trade secret)
Action keywords (password reset, account verification)
Financial keywords (wire transfer, payment required)
Personal information (SSN, credit card, date of birth)

How It Works

Function Name Analysis: Checks function names for suspicious patterns
Result Content Analysis: Analyzes function results for malicious content
Pattern Matching: Uses trained models to identify attack patterns
Threshold-Based Decisions:
- Score < input_safe_threshold (0.1) → PASS
- Score > input_block_threshold (0.5) → BLOCK
- Score between thresholds → Continue to Stage 2

Enabling Keyword Detection

Basic Setup

Enable keyword detection when analyzing a function:

from hipocap import Hipocap, observe

client = Hipocap.initialize(...)

@observe()
def process_user_data(user_id: str):
    data = fetch_user_data(user_id)
    
    result = client.analyze(
        function_name="process_user_data",
        function_result=data,
        function_args={"user_id": user_id},
        enable_keyword_detection=True  # Enable keyword detection
    )
    
    if not result.get("safe_to_use"):
        return {"error": "Sensitive keywords detected"}
    
    return data

Custom Keywords

Provide your own list of sensitive keywords:

result = client.analyze(
    function_name="process_user_data",
    function_result=data,
    function_args={"user_id": user_id},
    enable_keyword_detection=True,
    keywords=[
        "confidential",
        "classified",
        "top secret",
        "password reset",
        "account verification"
    ]
)

Default Keyword Patterns

HipoCap automatically detects common sensitive keyword patterns:

Security Keywords: confidential, classified, top secret, restricted, sensitive
Business Keywords: proprietary, trade secret, do not share
Action Keywords: password reset, account verification, urgent action
Financial Keywords: wire transfer, payment required, refund, account suspended
Personal Keywords: SSN, social security, credit card, date of birth, mother’s maiden name

Configuring Thresholds

You can adjust detection sensitivity by modifying thresholds in your policy:

result = client.analyze(
    function_name="get_user_data",
    function_result=user_data,
    function_args={"user_id": user_id},
    user_query=user_query,
    user_role="user",
    input_analysis=True,
    policy_key="default"  # Policy contains threshold settings
)

Thresholds are configured in the policy’s decision_thresholds:

input_safe_threshold: Score below this passes Stage 1 (default: 0.1)
input_block_threshold: Score above this blocks at Stage 1 (default: 0.5)
quarantine_safe_threshold: Score below this passes Stage 3 (default: 0.1)
quarantine_block_threshold: Score above this blocks at Stage 3 (default: 0.5)

Response Format

When keywords are detected, the analysis response includes:

{
    "keyword_detection": {
        "detected_keywords": ["confidential", "SSN"],
        "security_keywords": ["confidential"],
        "personal_keywords": ["SSN"],
        "keyword_positions": {
            "confidential": 3,  # Number of occurrences
            "SSN": 1
        }
    },
    "final_decision": "BLOCKED",
    "reason": "Sensitive keywords detected in function output",
    "input_score": 0.7,  # Risk score from Stage 1
    "safe_to_use": False
}

Practical Example: Email Processing

Here’s a complete example showing keyword detection in action:

from hipocap import Hipocap, observe

client = Hipocap.initialize(...)

@observe()
def process_email(email_id: str):
    email_content = fetch_email(email_id)
    
    result = client.analyze(
        function_name="process_email",
        function_result=email_content,
        function_args={"email_id": email_id},
        enable_keyword_detection=True,
        keywords=["confidential", "password reset", "account verification"],
        llm_analysis=True  # Also check for sensitive keywords in LLM analysis
    )
    
    # Check for detected keywords
    keyword_detection = result.get("keyword_detection", {})
    if keyword_detection and keyword_detection.get("detected_keywords"):
        # Sensitive keywords detected
        personal_keywords = keyword_detection.get("personal_keywords", [])
        if "SSN" in personal_keywords:
            # Block or redact sensitive data
            return redact_sensitive_data(email_content)
    
    # Check overall safety
    if not result.get("safe_to_use"):
        return {"error": "Content blocked", "reason": result.get("reason")}
    
    return email_content

Best Practices

Enable for Sensitive Functions - Always enable keyword detection for functions that handle sensitive data
Custom Keywords - Add domain-specific keywords relevant to your use case
Combine with Other Analysis - Use keyword detection alongside LLM and quarantine analysis for comprehensive protection
Adjust Thresholds - Fine-tune thresholds based on your false positive/negative rates
Monitor Results - Regularly review detected keywords to improve your keyword lists

Integration with Policies

Keyword detection can be configured in your governance policies:

{
  "functions": {
    "process_user_data": {
      "enable_keyword_detection": true,
      "keywords": ["SSN", "credit card", "date of birth"],
      "keyword_action": "BLOCK"
    }
  },
  "decision_thresholds": {
    "input_safe_threshold": 0.1,
    "input_block_threshold": 0.5
  }
}

Next Steps

Prompt Injection Protection - Learn about multi-stage analysis
Threat Categories - Complete threat reference
Setting up the Shield - Configure security analysis

Introduction

AI Security

Governance & RBAC

Observability

What is Keyword Detection?

How It Works

Enabling Keyword Detection

Basic Setup

Custom Keywords

Default Keyword Patterns

Configuring Thresholds

Response Format

Practical Example: Email Processing

Best Practices

Integration with Policies

Next Steps

​What is Keyword Detection?

​How It Works

​Enabling Keyword Detection

​Basic Setup

​Custom Keywords

​Default Keyword Patterns

​Configuring Thresholds

​Response Format

​Practical Example: Email Processing

​Best Practices

​Integration with Policies

​Next Steps

What is Keyword Detection?

How It Works

Enabling Keyword Detection

Basic Setup

Custom Keywords

Default Keyword Patterns

Configuring Thresholds

Response Format

Practical Example: Email Processing

Best Practices

Integration with Policies

Next Steps