Security

Security practices and vulnerability reporting for Sentinel.

Reporting Vulnerabilities

For security vulnerabilities, please contact maintainers directly instead of opening public issues.

See the Contributing Guide for contact details.

What to Include

1. Description of the vulnerability

2. Steps to reproduce

3. Potential impact

4. Any suggested fixes (optional)

Response Timeline

• Acknowledgment: Within 48 hours
• Initial assessment: Within 7 days
• Resolution timeline: Depends on severity

Responsible Disclosure

We follow responsible disclosure practices:

1. Report vulnerabilities privately

2. Allow time for a fix before public disclosure

3. Coordinate disclosure timing

Security Design

THSP Protocol

The THSP (Truth, Harm, Scope, Purpose) protocol provides defense-in-depth:

4-Layer Architecture

Memory Protection

Memory Shield provides HMAC-based signing to protect against memory injection attacks.

from sentinelseed import MemoryIntegrityChecker

checker = MemoryIntegrityChecker(secret_key="...")
signed = checker.sign_entry(entry)
result = checker.verify_entry(signed)

Database Protection

Database Guard prevents SQL injection and data exfiltration.

from sentinelseed import DatabaseGuard

guard = DatabaseGuard(
    blocked_tables=["users_private", "api_keys"],
)
result = guard.validate("SELECT * FROM users")

Security Best Practices

API Keys

• Never commit API keys to version control
• Use environment variables
• Rotate keys regularly

# Good
import os
api_key = os.environ.get("OPENAI_API_KEY")

# Bad
api_key = "sk-..."  # Never hardcode

Fail-Closed Mode

For high-stakes applications, use fail-closed mode:

from sentinelseed import LayeredValidator, ValidationConfig

config = ValidationConfig(
    fail_closed=True,  # Block on errors
)
validator = LayeredValidator(config=config)

Input Validation

Always validate user input before processing:

from sentinelseed import Sentinel

sentinel = Sentinel()
result = sentinel.validate_request(user_input)

if not result["should_proceed"]:
    # Block the request
    pass

Output Validation

Validate LLM outputs before returning to users:

from sentinelseed.detection import OutputValidator

validator = OutputValidator()
result = validator.validate(
    output_text=llm_response,
    input_text=user_input,
)

if not result.is_safe:
    # Block or sanitize the output
    pass

Attack Detection

Sentinel detects multiple attack vectors:

Prompt Injection

• Direct injection patterns
• Instruction override attempts
• Role-playing attacks
• Context manipulation

Jailbreak Attempts

• Multiple jailbreak detection patterns
• DAN, GPT-4, and other known jailbreaks
• Obfuscation detection

Data Exfiltration

• Sensitive data patterns
• PII detection
• API key/credential detection

Memory Attacks

• Memory injection patterns
• Context poisoning
• Trust score manipulation

Compliance

Sentinel supports compliance with:

• EU AI Act (Regulation 2024/1689)
• OWASP LLM Top 10
• OWASP Agentic Top 10 (2026)
• CSA AI Controls Matrix

See the Compliance Guide for details.

Data Privacy

No Data Storage

Sentinel does not store or log any validated content by default:

• No data sent to Sentinel servers
• All validation happens locally or through your own API keys
• History recording is opt-in and local

Local Processing

Heuristic validation is 100% local with no network calls.

Semantic validation uses your own API keys for OpenAI/Anthropic.

Dependencies

Sentinel has minimal dependencies to reduce supply chain risk:

Core (no external dependencies):

• Python 3.10+

Optional dependencies:

• openai for semantic validation
• anthropic for Claude-based validation
• Integration-specific packages

Audit Trail

For high-risk applications, implement logging:

import logging
from sentinelseed import LayeredValidator

logging.basicConfig(level=logging.INFO)

validator = LayeredValidator(
    config=ValidationConfig(
        log_validations=True,
    )
)

Updates

• Follow GitHub releases for security updates
• Subscribe to security advisories
• Update regularly to get latest detection patterns

Contact

• Security issues: Contact maintainers directly (see Contributing Guide)
• General questions: GitHub Issues
• Documentation: https://sentinelseed.dev/docs