No top_k in pipeline_kwargs gives a typeerror

[snackattas]

Describe the bug

"""
Bug: MaliciousURLs scanner crashes when using custom Model without top_k parameter

When providing a custom Model to MaliciousURLs scanner, if the pipeline_kwargs
don't include `top_k=None`, the scanner crashes with:

    TypeError: string indices must be integers, not 'str'

This happens because the HuggingFace classification pipeline returns different
formats based on whether top_k is specified:
- With top_k=None: Returns list of dicts [{"label": "...", "score": ...}]
- Without top_k: Returns just the top label as a string

The MaliciousURLs scanner expects the dict format but gets strings.

To reproduce:
    pip install llm-guard torch transformers
    python llm_guard_bug_repro.py

Environment:
    - Python: 3.12
    - llm-guard: 0.3.16
    - transformers: 4.x
"""

from llm_guard import scan_output
from llm_guard.model import Model
from llm_guard.output_scanners import MaliciousURLs

TEST_TEXT = """
Here is some information you requested.
Download the file here: http://example-malware-site.com/payload.exe
Let me know if you need anything else.
"""


def test_default_config():
    """Default config works - includes top_k=None internally."""
    print("=" * 60)
    print("TEST 1: Default MaliciousURLs() - WORKS")
    print("=" * 60)
    
    scanner = MaliciousURLs()
    run_scan(scanner)


def test_custom_model_no_top_k():
    """Custom Model without top_k - CRASHES."""
    print("\n" + "=" * 60)
    print("TEST 2: Custom Model without top_k - CRASHES")
    print("=" * 60)
    
    custom_model = Model(
        path="DunnBC22/codebert-base-Malicious_URLs",
        pipeline_kwargs={"max_length": 512, "truncation": True},
        # Missing: top_k=None
    )
    scanner = MaliciousURLs(model=custom_model)
    run_scan(scanner)


def test_custom_model_with_top_k():
    """Custom Model with top_k=None - WORKS."""
    print("\n" + "=" * 60)
    print("TEST 3: Custom Model with top_k=None - WORKS")
    print("=" * 60)
    
    custom_model = Model(
        path="DunnBC22/codebert-base-Malicious_URLs",
        pipeline_kwargs={"max_length": 512, "truncation": True, "top_k": None},
    )
    scanner = MaliciousURLs(model=custom_model)
    run_scan(scanner)


def run_scan(scanner):
    """Run the scanner and report result."""
    try:
        _, is_valid, risk_score = scan_output(
            scanners=[scanner],
            prompt="",
            output=TEST_TEXT
        )
        print(f"✓ SUCCESS - is_valid={is_valid}, risk_score={risk_score}")
    except TypeError as e:
        print(f"✗ CRASHED: {e}")


if __name__ == "__main__":
    test_default_config()
    test_custom_model_no_top_k()
    test_custom_model_with_top_k()
    
    print("\n" + "=" * 60)
    print("SUMMARY")
    print("=" * 60)
    print("""
The MaliciousURLs scanner crashes when using a custom Model if
pipeline_kwargs doesn't include `top_k=None`.

Suggested fix: MaliciousURLs scanner should either:
1. Document that top_k=None is required in custom Model configs
2. Automatically add top_k=None to pipeline_kwargs if not present
3. Handle both string and dict return formats from the pipeline
""")