FEAT: OWASP LLM10 cost amplification scenario with a token usage scorer
维护者通常 2 天内回复
还没有人认领这个 Issue。
评估
调研方向
Start by reading the existing RedTeamingAttack scenarios, scorer implementations, and how OpenAI and LiteLLM targets store token usage in prompt_metadata. Then check how FloatScaleThresholdScorer handles objective results and how target responses expose length and turn counts. Done means a cost amplification scenario and scorer measure token use against a baseline, provide the described fallback estimate, and enforce the proposed run limits.
由索引模型根据 Issue 内容生成。
描述
There is no scenario for OWASP LLM10 (unbounded consumption) in PyRIT today, and no open issue or PR for it. OpenAI and LiteLLM targets already save token usage in prompt_metadata, but no scorer reads it.
This is a measured risk. Prompt-only attacks raise output tokens many times over (arXiv:2512.23779, arXiv:2607.25936), and DOW-BENCH (arXiv:2609.28585) shows the cost grows across turns in agents.
I would like to add a CostAmplification scenario built on RedTeamingAttack. The adversarial chat stays polite and on topic, but asks for exhaustive answers, keeps asking for clarification, or expands every reply. A new TokenAmplificationScorer compares total tokens with the baseline run for the same objective. It scales the ratio to 0 to 1 against a cap, and FloatScaleThresholdScorer turns that into the objective result. HTTPTarget and PlaywrightTarget do not save tokens, so the scorer falls back to response length and turn count and marks the score as an estimate.
A self-ask scorer cannot see token counts, so it can only guess.
I saw the concern about expensive runs in #2533. So the scenario would be off by default, stop at a hard token budget without changing target settings, and print the maximum spend before it runs.
Would this fit PyRIT? If yes, I will share an implementation plan first.
- 主要语言
- Python
- 星标
- 4.6k
- 派生
- 962
- 平均合并
- 2 天 15 小时
- 30 天内合并 PR
- 390
环境准备
在浏览器里用你自己的 GitHub 账号启动这个项目的开发容器。
- 没有 Dockerfile 或 Docker Compose 文件
- 有 Pull Request 模板
- 没有贡献指南
从这里开始
- 先读完整个 Issue，再读项目的贡献指南。
- 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
- Fork 仓库，在一个分支上完成修改。
- 提交 Pull Request，并在描述里引用这个 Issue 编号。
microsoft/PyRIT 的其他 Issue
-
Bug: triage
难度 2/5 1-3 小时 新手友好度 72/100
维护者通常 2 天内回复
-
BUG: Video scorers block the event loop while decoding frames and audio可能已有人在做 @Shizoqua 今天认领。 未关闭
难度 2/5 1-3 小时 新手友好度 35/100
维护者通常 2 天内回复
-
BUG: HTTPTarget turns CRLF into LF in the request body可能已有人在做 @nniiovoo 今天认领。 未关闭
难度 2/5 1-3 小时 新手友好度 25/100
维护者通常 2 天内回复
-
BUG: multimodal router drops media metadata when forwarding seeds and responses可能已有人在做 @Shizoqua 今天认领。 未关闭Bug: triage
难度 3/5 1-2 天 新手友好度 35/100
维护者通常 2 天内回复
-
Bug: triage
难度 3/5 1-2 天 新手友好度 57/100
维护者通常 2 天内回复
相似的 Issue
-
bug documentation
难度 2/5 1-3 小时 新手友好度 78/100
awslabs/visual-asset-management-system#514 ·
维护者通常 1 天内回复
-
backend bug spec:REQ-053
难度 2/5 1-3 小时 新手友好度 66/100
nolte/kamerplanter#2547 ·
维护者通常 1 天内回复
-
难度 2/5 1-3 小时 新手友好度 75/100
EleutherAI/lm-evaluation-harness#4425 ·
维护者通常 1 天内回复
-
Gemini: respect PLANNING_DISABLED=1 in lifecycle context hooks可能已有人在做 @AmariAugustin 今天认领。 未关闭bug good first issue help wanted
难度 2/5 1-3 小时 新手友好度 87/100
OthmanAdi/planning-with-files#329 · 1 条评论 ·
维护者通常 1 天内回复
-
wr.dynamodb.read_items returns [{}] for a single partition value that does not exist可能已有人在做 @auraseb 今天认领。 未关闭
难度 2/5 1-3 小时 新手友好度 84/100
aws/aws-sdk-pandas#3520 ·
维护者通常 2 天内回复