FEAT: Add Echo Chamber multi-turn attack strategy
メンテナーはふだん 2 日以内に返信
まだ誰も着手していません。
評価
調査の方向性
Start with the existing Crescendo attack under pyrit/executor/attack/multi_turn/, since the issue reuses its turn loop, backtracking, AttackAdversarialConfig, AttackScoringConfig and SelfAskRefusalScorer. Confirm scope with maintainers first, since the issue's three questions are still open. Done means a new EchoChamberAttack class with its prompt under pyrit/datasets/executors/echo_chamber/, plus unit tests using mocked adversarial and objective targets that cover the seed phase, the echo turn, refusal with backtrack, and stop conditions.
索引モデルが issue の本文から書いたものです。
説明
Is your feature request related to a problem? Please describe.
PyRIT's multi-turn attacks (Crescendo, TAP, PAIR, RedTeaming) all escalate the request turn by turn. Echo Chamber (NeuralTrust, 2025) works differently. It gets the target to write benign text related to the objective, then keeps asking it to expand on its own earlier answers. The conversation drifts toward the objective through the model's own outputs, and no single user turn looks harmful, so per-turn filters and refusal checks don't catch it.
The paper reports 45.0% attack success on an AdvBench-derived set, against 28.6% for Crescendo and 9.5% for DAN on the same tasks. It covers a failure mode the current strategies don't.
I couldn't find an existing issue or PR for it.
Describe the solution you'd like
A new EchoChamberAttack in pyrit/executor/attack/multi_turn/, following the published flow:
- Seed. Derive neutral keywords from the objective, have the target write a few sentences from them, pick the one closest to the objective, and ask the target to expand on it.
- Echo loop. Each turn, the adversarial chat reads the transcript and writes a follow-up that builds on what the target already said, without stating the objective.
- Backtrack. On refusal, drop the exchange and retry from another angle.
- Score. Run the objective scorer each turn and stop on success or at
max_turns.
It would reuse what Crescendo uses: AttackAdversarialConfig, AttackScoringConfig, AttackConverterConfig, SelfAskRefusalScorer as the default refusal scorer, conversation-duplication backtracking, and MULTI_TURN + EDITABLE_HISTORY target requirements. The adversarial system prompt would go in pyrit/datasets/executors/echo_chamber/, adapted from NeuralTrust's MIT-licensed reference code with attribution.
The PR would include unit tests with mocked adversarial and objective targets (seed phase, echo turn, refusal and backtrack, stop conditions) and a doc entry next to the other multi-turn attacks.
Describe alternatives you've considered, if relevant
- Subclassing
CrescendoAttack. The turn loop and backtracking are the same, but the seed phase and the adversarial prompt differ, so I think a separate class is clearer. I'm fine going the other way if you prefer. - A converter. Wouldn't work, since the attack depends on what the target said in earlier turns.
Additional context
- NeuralTrust write-up (June 2025): https://neuraltrust.ai/research/echo-chamber
- Alobaid et al., "The Echo Chamber Multi-Turn LLM Jailbreak", arXiv:2601.05742: https://arxiv.org/abs/2601.05742
- Reference implementation (MIT): https://github.com/NeuralTrust/echo-chamber
The paper's numbers come from 12 objectives across GPT-4.1, GPT-5, and Gemini models, so the sample is small. I'll run it against a target I can reach and post the results here before opening the PR.
Questions:
- Is this in scope as a new multi-turn attack?
- Separate class or Crescendo subclass?
- Any preference for how the NeuralTrust attribution should look in the prompt YAML?
- 主要言語
- Python
- スター
- 4.6k
- フォーク
- 944
- 平均マージ
- 2日 18時間
- マージ済み PR（30日）
- 369
環境構築
このプロジェクトの開発コンテナを、あなたの GitHub アカウントでブラウザ上に起動します。
- Dockerfile・Docker Compose ファイルなし
- プルリクエストのテンプレートあり
- コントリビューションガイドなし
はじめの一歩
- issue を最後まで読み、次にプロジェクトのコントリビューションガイドを読みます。
- 着手することを issue にコメントします — 二人が同じ作業をするのを防げます。
- リポジトリをフォークし、ブランチを切って変更します。
- issue 番号を参照したプルリクエストを送ります。
microsoft/PyRIT のほかの issue
-
BUG: label queries only return the final conversation of an attack対応中かも @u7k4rs6 が 3 日前に担当しました。 オープン
難易度 2/5 1〜3時間 初心者へのやさしさ 85/100
microsoft/PyRIT#3095 · コメント 4 件 · 担当者 1 名 ·
メンテナーはふだん 2 日以内に返信
-
BUG Configuration keeps runtime-status errors after polling recovers対応中かも @joeyvictorino が 6 日前に担当しました。 オープンBug: triage GUI help wanted
難易度 2/5 1〜3時間 初心者へのやさしさ 86/100
microsoft/PyRIT#2868 · コメント 6 件 ·
メンテナーはふだん 2 日以内に返信
-
FEAT: Support PDF and file inputs via `binary_path` in OpenAI chat and response targets対応中かも @PranavD2905 が今日担当しました。 オープンfeature-request help wanted
難易度 4/5 3〜5日 初心者へのやさしさ 35/100
microsoft/PyRIT#3274 · コメント 1 件 ·
メンテナーはふだん 2 日以内に返信
-
feature-request
難易度 5/5 1週間以上 初心者へのやさしさ 18/100
メンテナーはふだん 2 日以内に返信
-
難易度 3/5 1〜2日 初心者へのやさしさ 35/100
メンテナーはふだん 2 日以内に返信
microsoft/PyRIT の issue をすべて見る
似ている issue
-
難易度 2/5 1〜3時間 初心者へのやさしさ 72/100
メンテナーはふだん 1 日以内に返信
-
難易度 2/5 1〜3時間 初心者へのやさしさ 76/100
omegaup/omegaup#10353 · コメント 1 件 ·
メンテナーはふだん 5 日以内に返信
-
bot-triaged module: LrScheduler module: optimizer triaged
難易度 2/5 1〜3時間 初心者へのやさしさ 70/100
メンテナーはふだん 1 日以内に返信
-
documentation glm5 llm-model
難易度 2/5 1〜3時間 初心者へのやさしさ 78/100
vllm-project/vllm-ascend#19062 ·
メンテナーはふだん 1 日以内に返信
-
author-page correction
難易度 2/5 1〜3時間 初心者へのやさしさ 68/100
acl-org/acl-anthology#10496 · コメント 1 件 ·
メンテナーはふだん 1 日以内に返信