Compliance Guide
How Sentinel helps meet regulatory requirements for AI systems.
Overview
Sentinel provides compliance checking for multiple regulatory frameworks:
| Framework | Module | Coverage |
|---|---|---|
| EU AI Act | EUAIActComplianceChecker | Strong |
| OWASP LLM Top 10 (2025) | OWASPLLMChecker | 60% Strong |
| OWASP Agentic Top 10 | OWASPAgenticChecker | 65% |
| CSA AI Controls Matrix | CSAAICMComplianceChecker | Partial |
EU AI Act
The EU AI Act (Regulation 2024/1689) establishes a risk-based regulatory framework for AI systems.
THSP Mapping
| EU AI Act Area | THSP Support | Coverage |
|---|---|---|
| Article 5: Prohibited Practices | Harm + Scope | Strong |
| Article 9: Risk Management | All Gates | Strong |
| Article 13: Transparency | Truth | Moderate |
| Article 14: Human Oversight | Scope + Purpose | Strong |
Usage
from sentinelseed.compliance import EUAIActComplianceChecker
checker = EUAIActComplianceChecker(api_key="sk-...")
result = checker.check_compliance(
content="Based on your social behavior, your credit score is...",
context="financial",
system_type="high_risk",
)
if not result.compliant:
print(f"Violations: {result.article_5_violations}")
print(f"Recommendations: {result.recommendations}")
Risk Levels
| Level | Description |
|---|---|
unacceptable | Banned practices detected |
high | Multiple gate failures |
limited | Single gate failure |
minimal | All gates pass |
Article 5 Detection
Sentinel's Harm and Scope gates detect prohibited practices:
- Subliminal manipulation
- Exploiting vulnerabilities
- Social scoring
- Emotion recognition misuse
# Detect subliminal manipulation
content = "Use hidden triggers to influence purchasing decisions"
result = checker.check_compliance(content)
# Returns: compliant=False, article_5_violations=["Article 5(1)(a): Potential manipulation"]
OWASP LLM Top 10 (2025)
The OWASP Top 10 for LLM Applications 2025 identifies critical risks in LLM applications.
THSP Coverage Summary
| Vulnerability | THSP Gate | Coverage |
|---|---|---|
| LLM01: Prompt Injection | Scope | Strong |
| LLM02: Sensitive Information Disclosure | Truth + Harm | Strong |
| LLM03: Supply Chain | — | Indirect |
| LLM04: Data and Model Poisoning | — | Indirect |
| LLM05: Improper Output Handling | Truth + Harm | Strong |
| LLM06: Excessive Agency | Scope + Purpose | Strong |
| LLM07: System Prompt Leakage | Scope | Moderate |
| LLM08: Vector and Embedding Weaknesses | — | Not Applicable |
| LLM09: Misinformation | Truth | Strong |
| LLM10: Unbounded Consumption | — | Not Applicable |
Usage
from sentinelseed.compliance import OWASPLLMChecker
checker = OWASPLLMChecker(api_key="sk-...")
# Check input (pre-inference)
input_result = checker.check_input(user_input)
if not input_result.secure:
print(f"Vulnerabilities: {input_result.vulnerabilities_detected}")
# Check output (post-inference)
output_result = checker.check_output(llm_output)
# Check full pipeline
result = checker.check_pipeline(
input_content=user_input,
output_content=llm_output,
)
Vulnerabilities Detected
| ID | Vulnerability | Detection Method |
|---|---|---|
| LLM01 | Prompt Injection | Scope gate, pattern matching |
| LLM02 | Sensitive Information Disclosure | Truth + Harm gates |
| LLM05 | Improper Output Handling | Output validation |
| LLM06 | Excessive Agency | Scope + Purpose gates |
| LLM07 | System Prompt Leakage | Scope gate |
| LLM09 | Misinformation | Truth gate |
OWASP Agentic Top 10 (2026)
Check coverage for agentic AI vulnerabilities.
from sentinelseed.compliance import OWASPAgenticChecker
checker = OWASPAgenticChecker()
result = checker.get_coverage_assessment()
print(f"Overall coverage: {result.overall_coverage}%")
for vuln in result.vulnerabilities:
print(f"{vuln.id}: {vuln.coverage_level}")
Coverage Summary
| ID | Vulnerability | Sentinel Coverage |
|---|---|---|
| AA01 | Agentic Misalignment | Full |
| AA02 | Tool Misuse | Full |
| AA03 | Privilege Escalation | Full |
| AA04 | Memory Poisoning | Full (Memory Shield) |
| AA05 | Identity Spoofing | Full |
| AA06 | Prompt/Instruction Theft | Partial |
| AA07 | Resource Overconsumption | Partial |
| AA08 | Multi-Agent Manipulation | Partial |
| AA09 | Supply Chain Vulnerabilities | None |
| AA10 | Insufficient Logging | None |
CSA AI Controls Matrix
Check against Cloud Security Alliance AI Controls Matrix.
from sentinelseed.compliance import CSAAICMComplianceChecker
checker = CSAAICMComplianceChecker(api_key="sk-...")
result = checker.check_compliance(
content=ai_output,
domains=["GRC", "SEC"], # Optional domain filter
)
Convenience Functions
from sentinelseed.compliance import (
check_eu_ai_act_compliance,
check_owasp_llm_compliance,
get_owasp_agentic_coverage,
check_csa_aicm_compliance,
)
# One-off checks
result = check_eu_ai_act_compliance(content, api_key="...")
result = check_owasp_llm_compliance(content, validation_type="output")
result = get_owasp_agentic_coverage()
result = check_csa_aicm_compliance(content, api_key="...")
Implementation Patterns
Pre-Deployment Compliance Check
from sentinelseed.compliance import EUAIActComplianceChecker
class ComplianceGuard:
def __init__(self):
self.checker = EUAIActComplianceChecker()
def check_before_deploy(self, ai_system_output: str) -> bool:
result = self.checker.check_compliance(
content=ai_system_output,
system_type="high_risk",
)
if not result.compliant:
self.log_violation(result)
return False
return True
Continuous Monitoring
class ComplianceMonitor:
def __init__(self):
self.checker = OWASPLLMChecker()
self.violations = []
def monitor_interaction(self, input_text: str, output_text: str):
result = self.checker.check_pipeline(
input_content=input_text,
output_content=output_text,
)
if not result.compliant:
self.violations.append({
"timestamp": datetime.now(),
"input": input_text,
"output": output_text,
"issues": result.issues,
})
Limitations
Sentinel compliance checking does NOT replace:
1. Legal compliance review
2. Technical documentation (Article 11)
3. Record-keeping systems (Article 12)
4. Conformity assessment procedures
5. Registration with EU database
6. Human oversight implementation
Use Sentinel as one layer of your compliance strategy, not the entire solution.
Timeline (EU AI Act)
| Date | Milestone | THSP Relevance |
|---|---|---|
| 2 Feb 2025 | Prohibited practices effective | THSP helps detect |
| 2 Aug 2025 | GPAI rules apply | Partial support |
| 2 Aug 2026 | High-risk rules apply | Strong support |