ruoxi_sun
58657935fc
Vendor LLMmap / llm-verify / llm-fingerprint-detector under bash/fingerprint/tools so the three fingerprint benchmarks run with only /data1/eval mounted (no /data1/xii dependency): - run.py DEFAULT_TOOLS_ROOT prefers builtin tools/, falls back to /data1/xii - exclude .git / node_modules / template backups - detector dist/ (pre-built) retained; node_modules not needed at runtime
111 lines
3.3 KiB
Python
111 lines
3.3 KiB
Python
"""Adversarial tests for fail-closed deep-analysis behavior."""
|
|
|
|
from types import SimpleNamespace
|
|
from unittest.mock import MagicMock
|
|
|
|
from src.schemas.analysis import ModelReport, RedFlag
|
|
from src.services.deep_analysis import (
|
|
DeepAnalysisService,
|
|
_evidence_quality,
|
|
_extract_identity_claims,
|
|
_extract_proxy_indicators,
|
|
_names_match,
|
|
_similarity_verdict,
|
|
)
|
|
|
|
|
|
def _service() -> DeepAnalysisService:
|
|
return DeepAnalysisService(MagicMock())
|
|
|
|
|
|
def _report(**overrides: object) -> ModelReport:
|
|
values: dict[str, object] = {
|
|
"model_name": "claude-sonnet-4-20250514",
|
|
"provider": "suspect",
|
|
"total_probes": 32,
|
|
"successful_probes": 32,
|
|
"evidence_quality": "SUFFICIENT",
|
|
}
|
|
values.update(overrides)
|
|
return ModelReport(**values)
|
|
|
|
|
|
def test_no_flags_is_not_called_legitimate() -> None:
|
|
verdict = _service()._determine_verdict([], [_report()])
|
|
assert verdict == "NO_FRAUD_SIGNALS"
|
|
|
|
|
|
def test_missing_evidence_never_receives_clean_verdict() -> None:
|
|
report = _report(
|
|
total_probes=10,
|
|
successful_probes=0,
|
|
errors=10,
|
|
error_rate=1.0,
|
|
evidence_quality="INSUFFICIENT",
|
|
)
|
|
flags = _service()._detect_red_flags([report], [])
|
|
verdict = _service()._determine_verdict(flags, [report])
|
|
|
|
assert any(flag.category == "evidence" for flag in flags)
|
|
assert verdict in {"SUSPICIOUS", "INCONCLUSIVE"}
|
|
|
|
|
|
def test_one_identity_mismatch_is_suspicious() -> None:
|
|
flags = [
|
|
RedFlag(
|
|
severity="HIGH",
|
|
category="identity",
|
|
description="Model family mismatch",
|
|
)
|
|
]
|
|
assert _service()._determine_verdict(flags, [_report()]) == "SUSPICIOUS"
|
|
|
|
|
|
def test_model_family_and_version_matching_is_strict() -> None:
|
|
assert _names_match("claude-sonnet-4-20250514", "claude-4")
|
|
assert not _names_match("claude-sonnet-4-20250514", "gpt-4o")
|
|
assert not _names_match("Opus 4.6", "claude-3.5-sonnet")
|
|
assert not _names_match("claude-3.5-sonnet", "claude-3.7-sonnet")
|
|
|
|
|
|
def test_identity_extraction_ignores_comparison_mentions() -> None:
|
|
results = [
|
|
SimpleNamespace(
|
|
response_text="I am Claude-4. I am not GPT-4 and should not be confused with it.",
|
|
prompt_category="identity",
|
|
)
|
|
]
|
|
assert _extract_identity_claims(results) == ["claude-4"]
|
|
|
|
|
|
def test_identity_extraction_supports_named_claude_variants() -> None:
|
|
results = [
|
|
SimpleNamespace(
|
|
response_text="My model name is Claude Sonnet 4.5.",
|
|
prompt_category="identity",
|
|
)
|
|
]
|
|
assert _extract_identity_claims(results) == ["claude sonnet 4.5"]
|
|
|
|
|
|
def test_proxy_disclosures_are_extracted() -> None:
|
|
results = [
|
|
SimpleNamespace(
|
|
response_text="Requests reach me through a managed proxy relay operated upstream."
|
|
)
|
|
]
|
|
indicators = _extract_proxy_indicators(results)
|
|
assert indicators
|
|
assert "proxy" in indicators[0]
|
|
|
|
|
|
def test_evidence_quality_thresholds() -> None:
|
|
assert _evidence_quality(8, 0.80) == "SUFFICIENT"
|
|
assert _evidence_quality(4, 0.60) == "DEGRADED"
|
|
assert _evidence_quality(2, 1.00) == "INSUFFICIENT"
|
|
|
|
|
|
def test_similarity_requires_enough_successful_probes() -> None:
|
|
assert _similarity_verdict(0.99, 3, 3) == "INCONCLUSIVE"
|
|
assert _similarity_verdict(0.91, 8, 8) == "SAME_MODEL"
|