DeepTeam is a framework to red team LLMs and AI agents.
import os
import uuid
import pytest
from deepeval.metrics.utils import initialize_model
from deepeval.tracing.types import AgentSpan, LlmSpan, Trace, TraceSpanStatus
from deepteam.trace_scanner.trace_scanner import TraceScanner
from deepteam.vulnerabilities.bias.bias import Bias
from deepteam.vulnerabilities.bias.template import BiasTemplate
def load_trace_fixture(filename: str) -> Trace:
test_dir = os.path.dirname(__file__)
filepath = os.path.join(test_dir, "traces", f"{filename}.json")
with open(filepath, "r") as f:
return Trace.model_validate_json(f.read())
@pytest.fixture(scope="module")
def evaluation_model():
model, using_native = initialize_model("gpt-4o")
return model, using_native
@pytest.fixture
def bias_vulnerability(evaluation_model):
model, _ = evaluation_model
return Bias(
evaluation_model=model,
async_mode=False
)
@pytest.fixture
def batch_evaluator_factory(evaluation_model):
model, using_native = evaluation_model
def _factory(limit: int) -> TraceScanner:
return TraceScanner(
model=model,
template=BiasTemplate,
limit=limit,
)
return _factory
class TestTraceScanner:
def test_sync_materialized_bias_finding(self, bias_vulnerability):
trace = load_trace_fixture("unsafe_bias_trace")
findings = bias_vulnerability._assess_trace(trace)
assert bias_vulnerability.vulnerable is True
assert len(findings) >= 1
assert any(f.vulnerabilityType == "gender" and f.outcome == "materialized" for f in findings)
def test_guardrail_mitigation_no_root_breach(self, bias_vulnerability):
trace = load_trace_fixture("guardrail_mitigated_trace")
findings = bias_vulnerability._assess_trace(trace)
assert len(findings) > 0
assert all(f.outcome == "mitigated" for f in findings)
assert bias_vulnerability.vulnerable is False
@pytest.mark.asyncio
async def test_async_deep_trace_empty_findings(self, bias_vulnerability):
trace = load_trace_fixture("safe_trace")
findings = await bias_vulnerability._a_assess_trace(trace)
assert findings == []
assert bias_vulnerability.vulnerable is False
def test_io_collapsing_wrapper_edge_case(self, batch_evaluator_factory):
trace_uuid = str(uuid.uuid4())
agent_uuid = str(uuid.uuid4())
llm_uuid = str(uuid.uuid4())
shared_input = "Write a controversial opinion."
shared_output = "Here is a highly controversial and biased opinion."
llm_span = LlmSpan(
uuid=llm_uuid,
name="actual_generation",
status=TraceSpanStatus.SUCCESS,
trace_uuid=trace_uuid,
parent_uuid=agent_uuid,
start_time=1_700_000_000.0,
end_time=1_700_000_001.0,
input=shared_input,
output=shared_output
)
agent_span = AgentSpan(
uuid=agent_uuid,
name="wrapper_agent",
status=TraceSpanStatus.SUCCESS,
trace_uuid=trace_uuid,
parent_uuid=None,
start_time=1_700_000_000.0,
end_time=1_700_000_001.0,
children=[llm_span],
input=shared_input,
output=shared_output,
available_tools=[],
agent_handoffs=[]
)
evaluator = batch_evaluator_factory(limit=40_000)
extracted_parent = evaluator._extract_span_with_findings(agent_span, child_findings=[])
assert extracted_parent.spanUuid == agent_uuid
assert extracted_parent.input is None
assert extracted_parent.output is None