Secret Injection

# Secret Injection

Inject credentials at runtime, never passing them through the LLM.

## Category
Tool Security

## Context
Tools that require authentication to external services.

## Problem
Credentials in prompts risk leakage and security breaches.

## Solution
Inject secrets server-side:
- **Context object**: Credentials in tool context
- **Never in params**: Don't accept secrets as input
- **Automatic injection**: Framework handles it
- **Secure storage**: Credentials encrypted at rest

## Example (Python)
```python
# BAD: Secret as parameter (exposed to LLM)
@tool
def call_api(api_key: str, endpoint: str) -> Result:
    return requests.get(endpoint, headers={"Authorization": api_key})

# GOOD: Secret injected via context (hidden from LLM)
@tool
def call_api(context: Context, endpoint: str) -> Result:
    api_key = context.get_secret("API_KEY")  # Injected, not passed
    return requests.get(endpoint, headers={"Authorization": api_key})
```

## Considerations
- Never accept credentials as tool parameters
- Use secure context injection
- Log access but never log credentials

## Related Patterns
- permission-gate
- identity-anchor
- audit-trail

---
Source: https://arcade.dev/patterns/secret-injection