AI Security Review

Review AI applications for OWASP LLM Top 10 and security best practices.

When to Use

• Reviewing AI application code before production
• Auditing for prompt injection vulnerabilities
• Checking data leakage and access control
• Validating OWASP LLM Top 10 compliance

OWASP LLM Top 10 Checklist

Prompt Injection Defense

def sanitize_user_input(text: str) -> str:
    """Remove common prompt injection patterns."""
    import re
    patterns = [
        r"(?i)ignore\s+(previous|above|all)\s+(instructions?|rules?)",
        r"(?i)you\s+are\s+now\s+(a|an)\s+",
        r"(?i)system\s*:\s*",
        r"(?i)###\s*(system|instruction|prompt)",
    ]
    for pattern in patterns:
        text = re.sub(pattern, "[FILTERED]", text)
    return text

# Always validate output before returning to user
def validate_output(response: str, system_prompt: str) -> str:
    """Ensure output doesn't leak system prompt."""
    if system_prompt[:50].lower() in response.lower():
        return "I cannot provide that information."
    return response

Security Review Checklist

## AI Security Review — [Service Name]

### Authentication & Authorization
- [ ] All endpoints require authentication
- [ ] Managed Identity used (no API keys in code)
- [ ] RBAC roles are least-privilege

### Input Validation
- [ ] User input sanitized before LLM
- [ ] File uploads validated (type, size)
- [ ] No SQL/command injection vectors

### Data Protection
- [ ] PII stripped before sending to LLM
- [ ] No secrets in prompts or context
- [ ] Audit logging captures access events

### Output Safety
- [ ] Content safety filter enabled
- [ ] Output validated before rendering
- [ ] System prompt not leakable

### Rate Limiting
- [ ] Per-user rate limits configured
- [ ] Token budget enforced
- [ ] Circuit breaker on LLM calls

Troubleshooting