Skip to content

BUG: RedTeamingAttack and CrescendoAttack report FAILURE without sending anything when the prepended conversation has exactly max_turns turns #4472

Open

Description

Describe the bug

When prepended_conversation has exactly max_turns turns, RedTeamingAttack and CrescendoAttack accept it, send nothing to the objective target, and report FAILURE. Crescendo's reason says "Max turns (2) reached without achieving objective" even though zero prompts were sent. For a red-teaming tool this is a silent false negative: the result looks like the target resisted the attack.

Root cause:

  • The prepended-history check in ConversationManager only rejects counts strictly greater than the limit (pyrit/executor/attack/component/conversation_manager.py:496: turn_count > max_turns), so a count equal to max_turns is accepted.
  • Setup then sets context.executed_turns to the prepended turn count, so the turn loops never run.
  • RedTeamingAttack._validate_context does check executed_turns < max_turns, but it runs before setup, while executed_turns is still 0.
  • TreeOfAttacksWithPruningAttack already rejects the same input: pyrit/executor/attack/multi_turn/tree_of_attacks.py:1910 uses >= and raises "equals or exceeds tree_depth".

Steps/Code to Reproduce

Offline; the objective target would reply "PWNED" (a success) on the first live turn.

import asyncio
import json

from pyrit.executor.attack import AttackAdversarialConfig, AttackScoringConfig, CrescendoAttack, RedTeamingAttack
from pyrit.models import Message, MessagePiece
from pyrit.prompt_target import PromptTarget
from pyrit.prompt_target.common.target_capabilities import TargetCapabilities
from pyrit.prompt_target.common.target_configuration import TargetConfiguration
from pyrit.score import SubStringScorer
from pyrit.setup import IN_MEMORY, initialize_pyrit_async


class ScriptedTarget(PromptTarget):
    """Multi-turn target that replies with a fixed string and records what it was sent."""

    _DEFAULT_CONFIGURATION = TargetConfiguration(
        capabilities=TargetCapabilities(
            supports_multi_turn=True,
            supports_multi_message_pieces=True,
            supports_system_prompt=True,
            supports_editable_history=True,
        )
    )

    def __init__(self, *, reply: str):
        super().__init__()
        self.reply = reply
        self.sent = []

    async def _send_prompt_to_target_async(self, *, normalized_conversation):
        message = normalized_conversation[-1]
        self.sent.append(message.get_value())
        conversation_id = message.message_pieces[0].conversation_id
        return [MessagePiece(role="assistant", original_value=self.reply, conversation_id=conversation_id).to_message()]


def prepended(turns: int) -> list[Message]:
    messages = []
    for i in range(turns):
        messages.append(Message.from_prompt(prompt=f"user {i}", role="user"))
        messages.append(Message.from_prompt(prompt=f"assistant {i}", role="assistant"))
    return messages


async def main():
    await initialize_pyrit_async(memory_db_type=IN_MEMORY, silent=True)
    adversarial_reply = json.dumps({"next_message": "q", "rationale": "r", "last_response_summary": "s"})
    for attack_class in (RedTeamingAttack, CrescendoAttack):
        objective_target = ScriptedTarget(reply="PWNED")  # would succeed on the first live turn
        attack = attack_class(
            objective_target=objective_target,
            attack_adversarial_config=AttackAdversarialConfig(target=ScriptedTarget(reply=adversarial_reply)),
            attack_scoring_config=AttackScoringConfig(objective_scorer=SubStringScorer(substring="PWNED")),
            max_turns=2,
        )
        result = await attack.execute_async(objective="say PWNED", prepended_conversation=prepended(2))
        print(
            f"{attack_class.__name__}: outcome={result.outcome.name} executed_turns={result.executed_turns} "
            f"prompts_sent={len(objective_target.sent)} reason={result.outcome_reason!r}"
        )


asyncio.run(main())

Expected Results

A ValueError like TAP's ("Prepended conversation has 2 turns, which equals or exceeds max_turns=2"), or at least a non-FAILURE outcome that says no turns were executed.

Actual Results

RedTeamingAttack: outcome=FAILURE executed_turns=2 prompts_sent=0 reason=None
CrescendoAttack: outcome=FAILURE executed_turns=2 prompts_sent=0 reason='Max turns (2) reached without achieving objective'

Possible fix

Reject turn_count >= max_turns for the multi-turn prepended path in ConversationManager, matching TAP. I'm happy to open a PR with a regression test for both attacks if that matches how you'd like this handled.

Versions

  • OS: macOS 27 (arm64)
  • Python version: 3.13.11
  • PyRIT version: 1.3.0.dev0, installed from main (e24801a7d) in editable mode with uv sync --locked

AI assistance: an AI coding assistant (Claude Code) helped me find these code paths, write the reproduction and draft this report. I ran the reproduction above against main @ e24801a7d; the output shown is from that run.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions