Skip to content

DeepTeam is a framework to red team LLMs and AI agents.

updated multi-turn to track turn_level_attack

A-Vamshi6e38b0fe0ca9
diff --git a/deepteam/attacks/multi_turn/bad_likert_judge/bad_likert_judge.py b/deepteam/attacks/multi_turn/bad_likert_judge/bad_likert_judge.py
index 720f8bdc..a888e9ac 100644
--- a/deepteam/attacks/multi_turn/bad_likert_judge/bad_likert_judge.py
+++ b/deepteam/attacks/multi_turn/bad_likert_judge/bad_likert_judge.py
@@ -144,13 +144,18 @@ class BadLikertJudge(BaseAttack):
# Randomly enhancing a turn attack
if self.turn_level_attacks and random.random() < 0.5:
attack = random.choice(self.turn_level_attacks)
+ turn_level_attack = attack
enhanced_attack = enhance_attack(
attack, enhanced_attack, self.simulator_model
)
assistant_response = model_callback(current_attack, turns)
turns.append(RTTurn(role="user", content=current_attack))
- turns.append(RTTurn(role="assistant", content=assistant_response))
+ if turn_level_attack is not None:
+ turns.append(RTTurn(role="assistant", content=assistant_response, turn_level_attack=turn_level_attack.get_name()))
+ else:
+ turns.append(RTTurn(role="assistant", content=assistant_response))
+ turn_level_attack = None
pbar.update(1)
@@ -252,13 +257,18 @@ class BadLikertJudge(BaseAttack):
# Randomly enhancing a turn attack
if self.turn_level_attacks and random.random() < 0.5:
attack = random.choice(self.turn_level_attacks)
+ turn_level_attack = attack
enhanced_attack = await a_enhance_attack(
attack, enhanced_attack, self.simulator_model
)
assistant_response = await model_callback(current_attack, turns)
turns.append(RTTurn(role="user", content=current_attack))
- turns.append(RTTurn(role="assistant", content=assistant_response))
+ if turn_level_attack is not None:
+ turns.append(RTTurn(role="assistant", content=assistant_response, turn_level_attack=turn_level_attack.get_name()))
+ else:
+ turns.append(RTTurn(role="assistant", content=assistant_response))
+ turn_level_attack = None
pbar.update(1)
diff --git a/deepteam/attacks/multi_turn/crescendo_jailbreaking/crescendo_jailbreaking.py b/deepteam/attacks/multi_turn/crescendo_jailbreaking/crescendo_jailbreaking.py
index 76b7815d..e126231a 100644
--- a/deepteam/attacks/multi_turn/crescendo_jailbreaking/crescendo_jailbreaking.py
+++ b/deepteam/attacks/multi_turn/crescendo_jailbreaking/crescendo_jailbreaking.py
@@ -183,6 +183,7 @@ class CrescendoJailbreaking(BaseAttack):
# Randomly enhancing a turn attack
if self.turn_level_attacks and random.random() < 0.5:
attack = random.choice(self.turn_level_attacks)
+ turn_level_attack = attack
current_attack = enhance_attack(
attack, current_attack, self.simulator_model
)
@@ -192,7 +193,11 @@ class CrescendoJailbreaking(BaseAttack):
last_response = self.generate_target_response(current_attack)
pbar_rounds.update(1)
- turns.append(RTTurn(role="assistant", content=last_response))
+ if turn_level_attack is not None:
+ turns.append(RTTurn(role="assistant", content=assistant_response, turn_level_attack=turn_level_attack.get_name()))
+ else:
+ turns.append(RTTurn(role="assistant", content=assistant_response))
+ turn_level_attack = None
is_refusal, refusal_rationale = self.get_refusal_score(
last_response, current_attack, vulnerability_data
@@ -340,6 +345,7 @@ class CrescendoJailbreaking(BaseAttack):
# Randomly enhancing a turn attack
if self.turn_level_attacks and random.random() < 0.5:
attack = random.choice(self.turn_level_attacks)
+ turn_level_attack = attack
current_attack = await a_enhance_attack(
attack, current_attack, self.simulator_model
)
@@ -351,7 +357,11 @@ class CrescendoJailbreaking(BaseAttack):
)
pbar_rounds.update(1)
- turns.append(RTTurn(role="assistant", content=last_response))
+ if turn_level_attack is not None:
+ turns.append(RTTurn(role="assistant", content=assistant_response, turn_level_attack=turn_level_attack.get_name()))
+ else:
+ turns.append(RTTurn(role="assistant", content=assistant_response))
+ turn_level_attack = None
is_refusal, refusal_rationale = await self.a_get_refusal_score(
last_response, current_attack, vulnerability_data
diff --git a/deepteam/attacks/multi_turn/linear_jailbreaking/linear_jailbreaking.py b/deepteam/attacks/multi_turn/linear_jailbreaking/linear_jailbreaking.py
index 62585fee..9b0d5430 100644
--- a/deepteam/attacks/multi_turn/linear_jailbreaking/linear_jailbreaking.py
+++ b/deepteam/attacks/multi_turn/linear_jailbreaking/linear_jailbreaking.py
@@ -133,13 +133,18 @@ class LinearJailbreaking(BaseAttack):
# Randomly enhancing a turn attack
if self.turn_level_attacks and random.random() < 0.5:
attack = random.choice(self.turn_level_attacks)
+ turn_level_attack = attack
current_attack = enhance_attack(
attack, current_attack, self.simulator_model
)
target_response = model_callback(current_attack, turns)
turns.append(RTTurn(role="user", content=current_attack))
- turns.append(RTTurn(role="assistant", content=target_response))
+ if turn_level_attack is not None:
+ turns.append(RTTurn(role="assistant", content=target_response, turn_level_attack=turn_level_attack.get_name()))
+ else:
+ turns.append(RTTurn(role="assistant", content=target_response))
+ turn_level_attack = None
pbar.update(1)
@@ -229,13 +234,18 @@ class LinearJailbreaking(BaseAttack):
# Randomly enhancing a turn attack
if self.turn_level_attacks and random.random() < 0.5:
attack = random.choice(self.turn_level_attacks)
+ turn_level_attack = attack
current_attack = await a_enhance_attack(
attack, current_attack, self.simulator_model
)
target_response = await model_callback(current_attack, turns)
turns.append(RTTurn(role="user", content=current_attack))
- turns.append(RTTurn(role="assistant", content=target_response))
+ if turn_level_attack is not None:
+ turns.append(RTTurn(role="assistant", content=target_response, turn_level_attack=turn_level_attack.get_name()))
+ else:
+ turns.append(RTTurn(role="assistant", content=target_response))
+ turn_level_attack = None
pbar.update(1)
diff --git a/deepteam/attacks/multi_turn/sequential_break/sequential_break.py b/deepteam/attacks/multi_turn/sequential_break/sequential_break.py
index ed48ef15..37b0a6e3 100644
--- a/deepteam/attacks/multi_turn/sequential_break/sequential_break.py
+++ b/deepteam/attacks/multi_turn/sequential_break/sequential_break.py
@@ -161,6 +161,7 @@ class SequentialJailbreak(BaseAttack):
# Randomly enhancing a turn attack
if self.turn_level_attacks and random.random() < 0.5:
attack = random.choice(self.turn_level_attacks)
+ turn_level_attack = attack
enhanced_attack = enhance_attack(
attack, enhanced_attack, self.simulator_model
)
@@ -168,7 +169,12 @@ class SequentialJailbreak(BaseAttack):
turns.append(RTTurn(role="user", content=current_attack))
assistant_response = model_callback(current_attack)
- turns.append(RTTurn(role="assistant", content=assistant_response))
+
+ if turn_level_attack is not None:
+ turns.append(RTTurn(role="assistant", content=assistant_response, turn_level_attack=turn_level_attack.get_name()))
+ else:
+ turns.append(RTTurn(role="assistant", content=assistant_response))
+ turn_level_attack = None
non_refusal_prompt = SequentialBreakTemplate.non_refusal(
current_attack, assistant_response
@@ -329,6 +335,7 @@ class SequentialJailbreak(BaseAttack):
# Randomly enhancing a turn attack
if self.turn_level_attacks and random.random() < 0.5:
attack = random.choice(self.turn_level_attacks)
+ turn_level_attack = attack
enhanced_attack = await a_enhance_attack(
attack, enhanced_attack, self.simulator_model
)
@@ -336,7 +343,12 @@ class SequentialJailbreak(BaseAttack):
turns.append(RTTurn(role="user", content=current_attack))
assistant_response = await model_callback(current_attack)
- turns.append(RTTurn(role="assistant", content=assistant_response))
+
+ if turn_level_attack is not None:
+ turns.append(RTTurn(role="assistant", content=assistant_response, turn_level_attack=turn_level_attack.get_name()))
+ else:
+ turns.append(RTTurn(role="assistant", content=assistant_response))
+ turn_level_attack = None
non_refusal_prompt = SequentialBreakTemplate.non_refusal(
current_attack, assistant_response
diff --git a/deepteam/attacks/multi_turn/tree_jailbreaking/tree_jailbreaking.py b/deepteam/attacks/multi_turn/tree_jailbreaking/tree_jailbreaking.py
index 7e008cee..aa4027e6 100644
--- a/deepteam/attacks/multi_turn/tree_jailbreaking/tree_jailbreaking.py
+++ b/deepteam/attacks/multi_turn/tree_jailbreaking/tree_jailbreaking.py
@@ -37,12 +37,14 @@ class TreeNode:
depth: int,
conversation_history=None,
parent: Optional["TreeNode"] = None,
+ turn_level_attack: Optional[BaseAttack] = None,
):
self.prompt = prompt
self.score = score
self.depth = depth
self.children = []
self.parent = parent
+ self.turn_level_attack = turn_level_attack
self.conversation_history = conversation_history or []
@@ -142,6 +144,10 @@ class TreeJailbreaking(BaseAttack):
turns.append(RTTurn(role="user", content=node.prompt))
assistant_response = model_callback(node.prompt, turns)
turns.append(RTTurn(role="assistant", content=assistant_response))
+ if node.turn_level_attack is not None:
+ turns.append(RTTurn(role="assistant", content=assistant_response, turn_level_attack=node.turn_level_attack.get_name()))
+ else:
+ turns.append(RTTurn(role="assistant", content=assistant_response))
return turns
@@ -375,8 +381,10 @@ class TreeJailbreaking(BaseAttack):
enhanced_attack = res.prompt
# Randomly enhancing a turn attack
+ turn_level_attack = None
if self.turn_level_attacks and random.random() < 0.5:
attack = random.choice(self.turn_level_attacks)
+ turn_level_attack = attack
enhanced_attack = enhance_attack(
attack, enhanced_attack, self.simulator_model
)
@@ -416,6 +424,7 @@ class TreeJailbreaking(BaseAttack):
depth=node.depth + 1,
conversation_history=conversation_json,
parent=node,
+ turn_level_attack=turn_level_attack
)
node.children.append(child_node)
@@ -592,8 +601,10 @@ class TreeJailbreaking(BaseAttack):
enhanced_attack = res.prompt
# Randomly enhancing a turn attack
+ turn_level_attack = None
if self.turn_level_attacks and random.random() < 0.5:
attack = random.choice(self.turn_level_attacks)
+ turn_level_attack = attack
enhanced_attack = await a_enhance_attack(
attack, enhanced_attack, self.simulator_model
)
@@ -632,6 +643,7 @@ class TreeJailbreaking(BaseAttack):
depth=node.depth + 1,
conversation_history=conversation_json,
parent=node,
+ turn_level_attack=turn_level_attack
)
##################################################