AI Security: The Basics Everyone Misses

Every team focuses on prompt injection. Most miss the basics that actually matter in production.

The Forgotten Fundamentals

1. Input Validation

Before the AI even sees input:

def validate_input(user_input: str) -> bool:
    # Length check
    if len(user_input) > 4000:
        return False
    
    # Rate limiting
    if exceeded_rate_limit(user_id):
        return False
    
    # Basic content filtering
    if contains_prohibited_content(user_input):
        return False
    
    return True

Most prompt injection attempts never get to your prompt if you filter input properly.

2. Output Validation

Check what AI generates before showing users:

def validate_output(ai_response: str) -> bool:
    # PII detection
    if contains_pii(ai_response):
        return False
    
    # Code execution attempts
    if contains_code_execution(ai_response):
        return False
    
    # Malicious links
    if contains_suspicious_urls(ai_response):
        return False
    
    return True

3. API Key Management

Never, ever hardcode API keys:

# NO
api_key = "sk-..." 

# YES
api_key = os.getenv("AZURE_OPENAI_KEY")

# BETTER
from azure.identity import DefaultAzureCredential
credential = DefaultAzureCredential()

Use managed identities. Always.

4. Logging and Monitoring

Log everything, but redact PII:

def log_interaction(user_input: str, ai_output: str):
    logger.info({
        "user_input": redact_pii(user_input),
        "ai_output": redact_pii(ai_output),
        "timestamp": datetime.utcnow(),
        "user_id": get_user_id(),
        "model": "gpt-4o",
        "cost": calculate_cost(...)
    })

You’ll need these logs when something goes wrong.

5. Cost Controls

Set budget limits:

async def call_ai_with_budget(prompt: str):
    daily_spend = await get_today_spend()
    
    if daily_spend > DAILY_BUDGET:
        raise BudgetExceededError()
    
    return await ai_client.complete(prompt)

One misconfigured bot can cost thousands.

The Actual Threat Model

Prompt injection? Possible, but rare in practice.

API key leaks? Common. Costs real money.

Data exfiltration? Happens. Users paste sensitive data.

Excessive costs? Very common. Bots hammering your API.

Toxic output? Will happen eventually. Filter it.

Focus on the common threats, not the theoretical ones.

Minimum Security Checklist

• Input validation and sanitization
• Output content filtering
• API keys in env variables or managed identity
• Rate limiting per user
• Cost monitoring and alerting
• Comprehensive logging (with PII redaction)
• Regular security testing
• Incident response plan

The Reality

Most AI security incidents aren’t sophisticated attacks. They’re:

• Leaked API keys on GitHub
• Bots running up costs
• Users pasting PII into prompts
• Missing rate limits

Fix the basics. They prevent 95% of actual problems.

Prompt injection defenses? Sure, add them. But only after you’ve handled the fundamentals.