DeepTeam is a framework to red team LLMs and AI agents.
attack simulator and red teamer updated to use rich.progress
A-Vamshi13c1532675f0
diff --git a/deepteam/attacks/attack_simulator/attack_simulator.py b/deepteam/attacks/attack_simulator/attack_simulator.pyindex d62142a8..a981ba43 100644--- a/deepteam/attacks/attack_simulator/attack_simulator.py+++ b/deepteam/attacks/attack_simulator/attack_simulator.py@@ -1,6 +1,5 @@import randomimport asyncio-from tqdm import tqdmfrom pydantic import BaseModelfrom typing import List, Optional, Unionimport inspect@@ -10,6 +9,7 @@ from deepeval.metrics.utils import initialize_modelfrom deepeval.metrics.utils import initialize_modelfrom deepteam.attacks import BaseAttack+from deepteam.utils import create_progress, rich_track, add_pbar, update_pbarfrom deepteam.vulnerabilities import BaseVulnerabilityfrom deepteam.attacks.multi_turn.types import CallbackTypefrom deepteam.errors import ModelRefusalError@@ -60,30 +60,33 @@ class AttackSimulator:num_vulnerability_types = sum(len(v.get_types()) for v in vulnerabilities)- pbar = tqdm(- vulnerabilities,- desc=f"π₯ Generating {num_vulnerability_types * attacks_per_vulnerability_type} attacks (for {num_vulnerability_types} vulnerability types across {len(vulnerabilities)} vulnerability(s))",- )- for vulnerability in pbar:- if simulator_model is not None:- vulnerability.simulator_model = simulator_model- test_cases.extend(- self.simulate_baseline_attacks(- attacks_per_vulnerability_type=attacks_per_vulnerability_type,- vulnerability=vulnerability,- ignore_errors=ignore_errors,- metadata=metadata,++ progress = create_progress()+ with progress:+ for vulnerability in rich_track(+ iterable=vulnerabilities,+ progress=progress,+ description=f"π₯ Generating {num_vulnerability_types * attacks_per_vulnerability_type} attacks (for {num_vulnerability_types} vulnerability types across {len(vulnerabilities)} vulnerability(s))"+ ):+ if simulator_model is not None:+ vulnerability.simulator_model = simulator_model+ test_cases.extend(+ self.simulate_baseline_attacks(+ attacks_per_vulnerability_type=attacks_per_vulnerability_type,+ vulnerability=vulnerability,+ ignore_errors=ignore_errors,+ metadata=metadata,+ ))- )if attacks is not None:# Enhance attacks by sampling from the provided distribution- pbar = tqdm(- test_cases,- desc=f"β¨ Simulating {num_vulnerability_types * attacks_per_vulnerability_type} attacks (using {len(attacks)} method(s))",- )attack_weights = [attack.weight for attack in attacks]- for test_case in pbar:+ for test_case in rich_track(+ iterable=test_cases,+ progress=progress,+ description=f"β¨ Simulating {num_vulnerability_types * attacks_per_vulnerability_type} attacks (using {len(attacks)} method(s))",+ ):# Randomly sample an enhancement based on the distributionsampled_attack = random.choices(attacks, weights=attack_weights, k=1@@ -114,70 +117,74 @@ class AttackSimulator:num_vulnerability_types = sum(len(v.get_types()) for v in vulnerabilities)- pbar = tqdm(- vulnerabilities,- desc=f"π₯ Generating {num_vulnerability_types * attacks_per_vulnerability_type} attacks (for {num_vulnerability_types} vulnerability types across {len(vulnerabilities)} vulnerability(s))",- )-- async def throttled_simulate_baseline_attack(vulnerability):- if simulator_model is not None:- vulnerability.simulator_model = simulator_model- async with semaphore: # Throttling applied here- result = await self.a_simulate_baseline_attacks(- attacks_per_vulnerability_type=attacks_per_vulnerability_type,- vulnerability=vulnerability,- ignore_errors=ignore_errors,- metadata=metadata,- )- pbar.update(1)- return result-- simulate_tasks = [- asyncio.create_task(- throttled_simulate_baseline_attack(vulnerability)- )- for vulnerability in vulnerabilities- ]+ progress = create_progress()+ with progress:- attack_results = await asyncio.gather(*simulate_tasks)- for result in attack_results:- test_cases.extend(result)- pbar.close()-- if attacks is not None:- # Enhance attacks by sampling from the provided distribution- pbar = tqdm(- total=len(test_cases),- desc=f"β¨ Simulating {num_vulnerability_types * attacks_per_vulnerability_type} attacks (using {len(attacks)} method(s))",+ task_id_vuln = add_pbar(+ progress,+ description=f"π₯ Generating {num_vulnerability_types * attacks_per_vulnerability_type} attacks (for {num_vulnerability_types} vulnerability types across {len(vulnerabilities)} vulnerability(s))",+ total=len(vulnerabilities))-- async def throttled_attack_method(- test_case: RTTestCase,- ):++ async def throttled_simulate_baseline_attack(vulnerability):+ if simulator_model is not None:+ vulnerability.simulator_model = simulator_modelasync with semaphore: # Throttling applied here- # Randomly sample an enhancement based on the distribution- if not attacks:- attack = BaselineAttack()- else:- attack_weights = [attack.weight for attack in attacks]- attack = random.choices(- attacks, weights=attack_weights, k=1- )[0]-- await self.a_enhance_attack(- attack=attack,- test_case=test_case,+ result = await self.a_simulate_baseline_attacks(+ attacks_per_vulnerability_type=attacks_per_vulnerability_type,+ vulnerability=vulnerability,ignore_errors=ignore_errors,+ metadata=metadata,)- pbar.update(1)+ update_pbar(progress, task_id_vuln)+ return result- await asyncio.gather(- *[- asyncio.create_task(throttled_attack_method(test_case))- for test_case in test_cases- ]- )- pbar.close()+ simulate_tasks = [+ asyncio.create_task(+ throttled_simulate_baseline_attack(vulnerability)+ )+ for vulnerability in vulnerabilities+ ]++ attack_results = await asyncio.gather(*simulate_tasks)+ for result in attack_results:+ test_cases.extend(result)++ if attacks is not None:+ # Enhance attacks by sampling from the provided distribution+ task_id_attack = add_pbar(+ progress,+ description=f"β¨ Simulating {num_vulnerability_types * attacks_per_vulnerability_type} attacks (using {len(attacks)} method(s))",+ total=len(test_cases)+ )+++ async def throttled_attack_method(+ test_case: RTTestCase,+ ):+ async with semaphore: # Throttling applied here+ # Randomly sample an enhancement based on the distribution+ if not attacks:+ attack = BaselineAttack()+ else:+ attack_weights = [attack.weight for attack in attacks]+ attack = random.choices(+ attacks, weights=attack_weights, k=1+ )[0]++ await self.a_enhance_attack(+ attack=attack,+ test_case=test_case,+ ignore_errors=ignore_errors,+ )+ update_pbar(progress, task_id_attack)++ await asyncio.gather(+ *[+ asyncio.create_task(throttled_attack_method(test_case))+ for test_case in test_cases+ ]+ )self.test_cases.extend(test_cases)return test_casesdiff --git a/deepteam/red_teamer/red_teamer.py b/deepteam/red_teamer/red_teamer.pyindex 89a0cf6c..12ea5158 100644--- a/deepteam/red_teamer/red_teamer.py+++ b/deepteam/red_teamer/red_teamer.py@@ -1,5 +1,4 @@import asyncio-from tqdm import tqdmfrom typing import Dict, List, Optional, Unionfrom rich.console import Consolefrom rich.table import Table@@ -19,7 +18,7 @@ from deepteam.frameworks.frameworks import AISafetyFrameworkfrom deepteam.frameworks.risk_category import RiskCategoryfrom deepteam.telemetry import capture_red_teamer_runfrom deepteam.attacks import BaseAttack-from deepteam.utils import validate_model_callback_signature+from deepteam.utils import validate_model_callback_signature, create_progress, rich_track, add_pbar, update_pbarfrom deepteam.vulnerabilities import BaseVulnerabilityfrom deepteam.vulnerabilities.types import VulnerabilityTypefrom deepteam.attacks.attack_simulator import AttackSimulator@@ -85,19 +84,21 @@ class RedTeamer:reuse_simulated_test_cases=reuse_simulated_test_cases,metadata=metadata,attacks_per_vulnerability_type=attacks_per_vulnerability_type,+ _print_assessment=False,)results = {}-- pbar = tqdm(- total=len(framework.risk_categories),- desc=f"β³ Running red-teaming for {framework.get_name()} Framework",- )- for risk_category in framework.risk_categories:- framework_assessment = assess_risk_category(risk_category)- results[risk_category.name] = framework_assessment- pbar.update(1)- pbar.close()+ progress = create_progress()+ with progress:+ task_id = add_pbar(+ progress,+ description=f"β³ Running red-teaming for {framework.get_name()} Framework",+ total=len(framework.risk_categories),+ )+ for risk_category in framework.risk_categories:+ framework_assessment = assess_risk_category(risk_category)+ results[risk_category.name] = framework_assessment+ update_pbar(progress, task_id)return results@@ -128,6 +129,7 @@ class RedTeamer:reuse_simulated_test_cases=reuse_simulated_test_cases,metadata=metadata,attacks_per_vulnerability_type=attacks_per_vulnerability_type,+ _print_assessment=False,)tasks = {@@ -137,18 +139,19 @@ class RedTeamer:results = {}- pbar = tqdm(- total=len(tasks),- desc=f"β³ Running red-teaming for {framework.get_name()} Framework",- )-- for task in asyncio.as_completed(tasks):- result = await task- name = tasks[task]- results[name] = result- pbar.update(1)+ progress = create_progress()+ with progress:+ task_id = add_pbar(+ progress,+ description=f"β³ Running red-teaming for {framework.get_name()} Framework",+ total=len(framework.risk_categories),+ )+ for task in asyncio.as_completed(tasks):+ result = await task+ name = tasks[task]+ results[name] = result+ update_pbar(progress, task_id)- pbar.close()return resultsdef red_team(@@ -163,6 +166,7 @@ class RedTeamer:ignore_errors: bool = False,reuse_simulated_test_cases: bool = False,metadata: Optional[dict] = None,+ _print_assessment: Optional[bool] = True,):if not framework and not vulnerabilities:raise ValueError(@@ -208,21 +212,20 @@ class RedTeamer:ignore_errors=ignore_errors,reuse_simulated_test_cases=reuse_simulated_test_cases,metadata=metadata,+ _print_assessment=_print_assessment))else:- if framework:- if framework._has_dataset:- pbar = tqdm(- range(framework.num_attacks),- desc=f"π₯ Fetching {framework.num_attacks} attacks from {framework.get_name()} Dataset",+ if framework and framework._has_dataset:+ progress = create_progress()+ with progress:+ task_id = add_pbar(+ progress,+ description=f"π₯ Fetching {framework.num_attacks} attacks from {framework.get_name()} Dataset",+ total=framework.num_attacks,)framework.load_dataset()- pbar.update(framework.num_attacks)- pbar.close()- else:- attacks = framework.attacks- vulnerabilities = framework.vulnerabilities+ update_pbar(progress, task_id, advance_to_end=True)assert not inspect.iscoroutinefunction(model_callback@@ -289,14 +292,16 @@ class RedTeamer:)if framework and framework._has_dataset:- pbar = tqdm(- total=len(simulated_test_cases),- desc=f"π Evaluating {len(simulated_test_cases)} test cases using {framework.get_name()} risk categories",- )- red_teaming_test_cases = framework.assess(- model_callback, pbar, ignore_errors- )- pbar.close()+ progress = create_progress()+ with progress:+ task_id = add_pbar(+ progress,+ description=f"π Evaluating {len(simulated_test_cases)} test cases using {framework.get_name()} risk categories",+ total=len(simulated_test_cases),+ )+ red_teaming_test_cases = framework.assess(+ model_callback, progress, task_id, ignore_errors+ )else:total_attacks = sum(len(test_cases)@@ -305,29 +310,30 @@ class RedTeamer:num_vulnerability_types = sum(len(v.get_types()) for v in vulnerabilities)- pbar = tqdm(- total=total_attacks,- desc=f"π Evaluating {num_vulnerability_types} vulnerability types across {len(vulnerabilities)} vulnerability(s)",- )+ progress = create_progress()+ with progress:+ task_id = add_pbar(+ progress,+ description=f"π Evaluating {num_vulnerability_types} vulnerability types across {len(vulnerabilities)} vulnerability(s)",+ total=total_attacks,+ )- red_teaming_test_cases: List[RTTestCase] = []+ red_teaming_test_cases: List[RTTestCase] = []- for (- vulnerability_type,- test_cases,- ) in vulnerability_type_to_attacks_map.items():- rt_test_cases = self._evaluate_vulnerability_type(- model_callback,- vulnerabilities,+ for (vulnerability_type,test_cases,- ignore_errors=ignore_errors,- )- red_teaming_test_cases.extend(rt_test_cases)-- pbar.update(len(test_cases))+ ) in vulnerability_type_to_attacks_map.items():+ rt_test_cases = self._evaluate_vulnerability_type(+ model_callback,+ vulnerabilities,+ vulnerability_type,+ test_cases,+ ignore_errors=ignore_errors,+ )+ red_teaming_test_cases.extend(rt_test_cases)- pbar.close()+ update_pbar(progress, task_id, advance=len(test_cases))self.risk_assessment = RiskAssessment(overview=construct_risk_assessment_overview(@@ -336,7 +342,9 @@ class RedTeamer:test_cases=red_teaming_test_cases,)self.test_cases = red_teaming_test_cases- self._print_risk_assessment()++ if _print_assessment:+ self._print_risk_assessment(self.risk_assessment)return self.risk_assessment@@ -352,6 +360,7 @@ class RedTeamer:ignore_errors: bool = False,reuse_simulated_test_cases: bool = False,metadata: Optional[dict] = None,+ _print_assessment: Optional[bool] = True,):if not framework and not vulnerabilities:raise ValueError(@@ -381,16 +390,15 @@ class RedTeamer:if framework:if framework._has_dataset:- pbar = tqdm(- range(framework.num_attacks),- desc=f"π₯ Fetching {framework.num_attacks} attacks from {framework.get_name()} Dataset",- )- framework.load_dataset()- pbar.update(framework.num_attacks)- pbar.close()- else:- attacks = framework.attacks- vulnerabilities = framework.vulnerabilities+ progress = create_progress()+ with progress:+ task_id = add_pbar(+ progress,+ description=f"π₯ Fetching {framework.num_attacks} attacks from {framework.get_name()} Dataset",+ total=framework.num_attacks,+ )+ framework.load_dataset()+ update_pbar(progress, task_id, advance_to_end=True)if evaluation_model is not None:self.evaluation_model = evaluation_model@@ -454,14 +462,16 @@ class RedTeamer:)if framework and framework._has_dataset:- pbar = tqdm(- total=len(simulated_test_cases),- desc=f"π Evaluating {len(simulated_test_cases)} test cases using {framework.get_name()} risk categories",- )- red_teaming_test_cases = await framework.a_assess(- model_callback, pbar, ignore_errors- )- pbar.close()+ progress = create_progress()+ with progress:+ task_id = add_pbar(+ progress,+ description=f"π Evaluating {len(simulated_test_cases)} test cases using {framework.get_name()} risk categories",+ total=len(simulated_test_cases),+ )+ red_teaming_test_cases = await framework.a_assess(+ model_callback, progress, task_id, ignore_errors+ )else:semaphore = asyncio.Semaphore(self.max_concurrent)total_attacks = sum(@@ -471,36 +481,38 @@ class RedTeamer:num_vulnerability_types = sum(len(v.get_types()) for v in vulnerabilities)- pbar = tqdm(- total=total_attacks,- desc=f"π Evaluating {num_vulnerability_types} vulnerability types across {len(vulnerabilities)} vulnerability(s)",- )-- red_teaming_test_cases: List[RTTestCase] = []+ progress = create_progress()+ with progress:+ task_id = add_pbar(+ progress,+ description=f"π Evaluating {num_vulnerability_types} vulnerability types across {len(vulnerabilities)} vulnerability(s)",+ total=total_attacks,+ )- async def throttled_evaluate_vulnerability_type(- vulnerability_type, attacks- ):- async with semaphore:- test_cases = await self._a_evaluate_vulnerability_type(- model_callback,- vulnerabilities,- vulnerability_type,- attacks,- ignore_errors=ignore_errors,- )- red_teaming_test_cases.extend(test_cases)- pbar.update(len(attacks))+ red_teaming_test_cases: List[RTTestCase] = []- # Create a list of tasks for evaluating each vulnerability, with throttling- tasks = [- throttled_evaluate_vulnerability_type(+ async def throttled_evaluate_vulnerability_type(vulnerability_type, attacks- )- for vulnerability_type, attacks in vulnerability_type_to_attacks_map.items()- ]- await asyncio.gather(*tasks)- pbar.close()+ ):+ async with semaphore:+ test_cases = await self._a_evaluate_vulnerability_type(+ model_callback,+ vulnerabilities,+ vulnerability_type,+ attacks,+ ignore_errors=ignore_errors,+ )+ red_teaming_test_cases.extend(test_cases)+ update_pbar(progress, task_id, advance=len(attacks))++ # Create a list of tasks for evaluating each vulnerability, with throttling+ tasks = [+ throttled_evaluate_vulnerability_type(+ vulnerability_type, attacks+ )+ for vulnerability_type, attacks in vulnerability_type_to_attacks_map.items()+ ]+ await asyncio.gather(*tasks)self.risk_assessment = RiskAssessment(overview=construct_risk_assessment_overview(@@ -509,7 +521,10 @@ class RedTeamer:test_cases=red_teaming_test_cases,)self.test_cases = red_teaming_test_cases- self._print_risk_assessment()++ if _print_assessment:+ self._print_risk_assessment(self.risk_assessment)+return self.risk_assessmentdef _attack(@@ -711,8 +726,8 @@ class RedTeamer:)return red_teaming_test_cases- def _print_risk_assessment(self):- if self.risk_assessment is None:+ def _print_risk_assessment(self, risk_assessment=None):+ if risk_assessment is None:returnconsole = Console()@@ -746,7 +761,7 @@ class RedTeamer:table.add_column("Status", justify="center", width=10)# Add rows- for case in self.risk_assessment.test_cases:+ for case in risk_assessment.test_cases:status = ("Passed"if case.score and case.score > 0@@ -789,13 +804,13 @@ class RedTeamer:console.print("\n" + "=" * 80)console.print(- f"[bold magenta]π DeepTeam Risk Assessment[/bold magenta] ({self.risk_assessment.overview.errored} errored)"+ f"[bold magenta]π DeepTeam Risk Assessment[/bold magenta] ({risk_assessment.overview.errored} errored)")console.print("=" * 80)# Sort vulnerability type results by pass rate in descending ordersorted_vulnerability_results = sorted(- self.risk_assessment.overview.vulnerability_type_results,+ risk_assessment.overview.vulnerability_type_results,key=lambda x: x.pass_rate,reverse=True,)@@ -821,7 +836,7 @@ class RedTeamer:# Sort attack method results by pass rate in descending ordersorted_attack_method_results = sorted(- self.risk_assessment.overview.attack_method_results,+ risk_assessment.overview.attack_method_results,key=lambda x: x.pass_rate,reverse=True,)@@ -851,6 +866,11 @@ class RedTeamer:console.print("=" * 80 + "\n")def _print_framework_overview_table(self, framework_results: dict):++ for category_name in sorted(framework_results.keys()):+ assessment = framework_results[category_name]+ self._print_risk_assessment(assessment)+console = Console()console.print("\n" + "=" * 80)