Skip to content

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

updated sequential

A-Vamshi73d54ffd7867
diff --git a/deepteam/attacks/multi_turn/sequential_break/sequential_break.py b/deepteam/attacks/multi_turn/sequential_break/sequential_break.py
index 7ddd860c..417a6776 100644
--- a/deepteam/attacks/multi_turn/sequential_break/sequential_break.py
+++ b/deepteam/attacks/multi_turn/sequential_break/sequential_break.py
@@ -77,18 +77,25 @@ class SequentialJailbreak(BaseAttack):
f"Vulnerability: {vulnerability} | Type: {vulnerability_type}"
)
+ # 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 original_attack is None:
- raise ValueError(
- "No user turn found in the conversation history to start the attack."
- )
+ if current_attack is None:
+ raise ValueError("No user turns found in conversation history")
+
+ # If the last turn is from user, we need a model response before simulation
+ if len(turns) == 0 or turns[-1].role == "user":
+ target_response = model_callback(current_attack, turns)
+ turns.append(RTTurn(role="assistant", content=target_response))
+ else:
+ target_response = turns[-1].content
- current_attack = original_attack
eval_flag = False
eval_percentage = 0
@@ -233,18 +240,25 @@ class SequentialJailbreak(BaseAttack):
if turns is None:
turns = []
+ # 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 original_attack is None:
- raise ValueError(
- "No user turn found in the conversation history to start the attack."
- )
+ if current_attack is None:
+ raise ValueError("No user turns found in conversation history")
+
+ # If last turn is user, generate a model response before the loop
+ if len(turns) == 0 or turns[-1].role == "user":
+ target_response = await model_callback(current_attack, turns)
+ turns.append(RTTurn(role="assistant", content=target_response))
+ else:
+ target_response = turns[-1].content
- current_attack = original_attack
eval_flag = False
eval_percentage = 0
@@ -320,6 +334,7 @@ class SequentialJailbreak(BaseAttack):
judge_res: QuestionBankJudge = await a_generate(
judge_prompt, QuestionBankJudge, self.simulator_model
)
+ eval_flag = judge_res.classification == "Harmful"
elif self.attack_type == "game_environment":
judge_prompt = SequentialBreakTemplate.game_environment_judge(
original_attack,
@@ -330,6 +345,7 @@ class SequentialJailbreak(BaseAttack):
judge_res: GameEnvironmentJudge = await a_generate(
judge_prompt, GameEnvironmentJudge, self.simulator_model
)
+ eval_flag = judge_res.classification == "Harmful"
else:
judge_prompt = SequentialBreakTemplate.dialogue_judge(
original_attack,
@@ -340,10 +356,11 @@ class SequentialJailbreak(BaseAttack):
judge_res: DialogueJudge = await a_generate(
judge_prompt, DialogueJudge, self.simulator_model
)
+ eval_flag = judge_res.classification == "Harmful"
pbar.update(1)
- if judge_res.classification == "Harmful":
+ if eval_flag == "Harmful":
pbar.close()
return turns