Skip to content

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

fixed jailbreaking attack

A-Vamshie5e3ef554873
diff --git a/deepteam/attacks/multi_turn/linear_jailbreaking/linear_jailbreaking.py b/deepteam/attacks/multi_turn/linear_jailbreaking/linear_jailbreaking.py
index a819bf3a..4e397ce0 100644
--- a/deepteam/attacks/multi_turn/linear_jailbreaking/linear_jailbreaking.py
+++ b/deepteam/attacks/multi_turn/linear_jailbreaking/linear_jailbreaking.py
@@ -160,27 +160,91 @@ class LinearJailbreaking(BaseAttack):
async def a_enhance(
self,
- attack: str,
model_callback: CallbackType,
+ turns: Optional[List[RTTurn]] = None,
simulator_model: Optional[Union[DeepEvalBaseLLM, str]] = None,
- turn_history: Optional[List[Turn]] = None,
- ) -> str:
+ ) -> List[RTTurn]:
+ if turns is None:
+ turns = []
+
self.simulator_model, _ = initialize_model(simulator_model)
pbar = tqdm(
- total=self.turns, desc="...... ⛓️ Linear Jailbreaking", leave=False
+ total=self.num_turns,
+ desc="...... ⛓️ Linear Jailbreaking",
+ leave=False,
)
- # Initialize conversation history for tracking attempts
- current_attack = attack
- turns = turn_history or []
+ # Get current attack from last user turn (simple reverse search)
+ current_attack = None
+ original_attack = None
+ for turn in reversed(turns):
+ if turn.role == "user":
+ current_attack = turn.content
+ original_attack = turn.content
+ break
+
+ if current_attack is None:
+ raise ValueError("No user turns found in conversation history")
+
+ # Simple logic based on last turn
+ if len(turns) > 0 and turns[-1].role == "assistant":
+ # Last turn is assistant - find the last user turn before it
+ user_turn_content = None
+ for turn in reversed(turns[:-1]): # Skip the last assistant turn
+ if turn.role == "user":
+ user_turn_content = turn.content
+ break
+
+ if user_turn_content is not None:
+ # We have a user turn before the assistant, generate score + improvement
+ target_response = turns[-1].content
+ current_attack = user_turn_content
+
+ judge_prompt = JailBreakingTemplate.linear_judge(
+ original_attack, current_attack, target_response
+ )
+ feedback: Feedback = await a_generate(
+ judge_prompt, Feedback, self.simulator_model
+ )
- for i in range(self.turns):
+ if feedback.jailbroken:
+ pbar.close()
+ return turns
+
+ improvement_prompt = JailBreakingTemplate.improvement_prompt(
+ turns, feedback.suggestion
+ )
+ improvement: Improvement = await a_generate(
+ improvement_prompt, Improvement, self.simulator_model
+ )
+ next_attack = improvement.new_prompt
+
+ non_refusal_prompt = BaseMultiTurnTemplate.non_refusal(
+ original_attack, next_attack
+ )
+ non_refusal_res: NonRefusal = await a_generate(
+ non_refusal_prompt, NonRefusal, self.simulator_model
+ )
+
+ if non_refusal_res.classification == "Refusal":
+ pbar.close()
+ raise ModelRefusalError(entity=self.get_name())
+
+ current_attack = next_attack
+ else:
+ # Last turn is user - call model callback first
target_response = await model_callback(current_attack, turns)
- turns = update_turn_history(turns, current_attack, target_response)
+ turns.append(RTTurn(role="assistant", content=target_response))
+
+ # Main simulation loop
+ for i in range(self.num_turns):
+ target_response = await model_callback(current_attack, turns)
+ turns.append(RTTurn(role="user", content=current_attack))
+ turns.append(RTTurn(role="assistant", content=target_response))
judge_prompt = JailBreakingTemplate.linear_judge(
- attack, current_attack, target_response
+ original_attack, current_attack, target_response
)
feedback: Feedback = await a_generate(
judge_prompt, Feedback, self.simulator_model
@@ -200,25 +264,21 @@ class LinearJailbreaking(BaseAttack):
next_attack = improvement.new_prompt
non_refusal_prompt = BaseMultiTurnTemplate.non_refusal(
- attack, next_attack
+ original_attack, next_attack
)
non_refusal_res: NonRefusal = await a_generate(
non_refusal_prompt, NonRefusal, self.simulator_model
)
- classification = non_refusal_res.classification
- if classification == "Refusal":
+
+ if non_refusal_res.classification == "Refusal":
pbar.n = pbar.total
pbar.close()
raise ModelRefusalError(entity=self.get_name())
- else:
- current_attack = next_attack
- if i == self.turns - 1:
- turns.append(Turn(role="user", content=current_attack))
+ current_attack = next_attack
pbar.update(1)
pbar.close()
-
return turns
def get_name(self) -> str: