LLM Excessive Agency

Description

Detects LLM implementations with excessive autonomy that can lead to unintended consequences. OWASP LLM08 - Excessive Agency. Excessive agency occurs when LLMs are granted: - Ability to perform destructive operations without confirmation - Auto-execution of LLM-generated code or commands - Direct database modifications without approval - Financial transactions without human oversight

What Shoulder detects

How to fix

Require human approval for sensitive operations and set iteration limits.

```python
SENSITIVE_OPERATIONS = {'delete_record', 'send_email', 'process_payment'}
MAX_ITERATIONS = 10

for tool_call in response.choices[0].message.tool_calls:
    if tool_call.function.name in SENSITIVE_OPERATIONS:
        await queue_for_approval(tool_call)
    else:
        await execute_safe_tool_call(tool_call)
```

Learn more: https://shoulder.dev/learn/python/cwe-862/llm-excessive-agency

Applies to

References

Scan for this issue

Related rules