/
划分小节”,但不能增删改内容,那是不是原文的逻辑分块?不对,等下原文是连续的?等下仔细看原文的内容: 第一段:Prompt injection is the dominant unsolved security risk for LLM agents, and the community now agrees that robustness must be measured against adaptive attackers rather than fixed benchmark injections. Many adaptive red teamers exist, and most are built to maximize attack success rate (ASR). Yet ASR counts how often a target breaks, not in how many different ways. A red teamer can achieve a high ASR by reusing a few techniques that work, leaving the others untested and absent from the attacks used to harden models. 这部分是背景和现有问题,可以作为一个小节?不对,但不能随便加标题文字啊,哦不对,不对,规则说不得增删改内容,所以h2/h3的文本必须是原文里的?不对,等下看规则2:“每个自然段落用
包裹,长文用
/
划分小节”,哦,不对,是不是如果原文有明显的小节逻辑,就划分,没有的话不用硬加?不对,再看原文: 然后接下来是:We present AutoPI, an adaptive red teamer designed to find many different ways to break a target, not just to break it often. AutoPI uses PIAtlas, a taxonomy of prompt-injection techniques, both to guide its search and to measure coverage. For each scenario, an attacker LLM writes an injection for every sub-technique in the taxonomy and then reworks the ones that fail using feedback from an LLM judge. The method is black-box and training-free, it plugs into existing benchmarks in place of their fixed attack templates, and it stays up to date as the taxonomy evolves. 这部分是介绍AutoPI的核心内容,可以作为一个小节?然后接下来是评估部分:We evaluate AutoPI on AgentDojo and IPI Arena against open-weight and proprietary target models. Alongside ASR, we report technique coverage, the fraction of taxonomy classes, techniques, and sub-techniques that yield at least one successful injection. 这部分是评估设置?然后是评估结果:On AgentDojo, AutoPI finds a successful injection against Qwen3.6-35B in all six classes of the taxonomy and for more than 90% of its techniques. Against Claude Haiku 5.5, the most robust target we evaluate, it achieves between 20% and 50% higher ASR and between 25% and 400% higher class-level coverage than recent state-of-the-art adaptive red teamers, while requiring no training. 最后是:We release our code and the generated attacks to support reproducible, coverage-aware evaluation. 不对,但是不能随便加h2的文本啊,因为规则说不能增删改内容。哦,不对,是不是我理解错了?等下看规则2,是“长文用
/
划分小节”,如果原文没有小标题的话,是不是不需要硬加?不对,再仔细看用户给的正文,是不是所有内容都是连续的?不对,等下,有没有可能用户希望把核心内容用strong标?比如“Prompt injection is the dominant unsolved security risk for LLM agents”这个可以加strong?然后“We present AutoPI”这个也可以加strong? 等下,先理清楚,首先,不能增删改任何内容,所以所有文本都必须是原文的,不能加h2的标题文字,因为原文里没有这些标题,加了就是增内容了。哦,对哦!规则4说保留原文所有信息,不得增删改内容,所以不能自己加小标题的文字,因为原文里没有。那h2/h3什么时候用?哦,不对,是不是原文里如果有小节的标识?不对,原文里没有,那是不是就不用h2/h3?不对,再看规则2:“每个自然段落用
包裹,长文用
/
划分小节”,哦,可能如果原文有逻辑分层,比如原文里的不同部分,比如背景、提出方法、评估、开源,但是不能用自己的话做标题,那怎么办?不对,等下,再看原文的内容,是不是本身就是几个自然段?不对,用户给的正文是一整段?不对,看用户给的正文: 正文如下: Prompt injection is the dominant unsolved security risk for LLM agents, and the community now agrees that robustness must be measured against adaptive attackers rather than fixed benchmark injections. Many adaptive red teamers exist, and most are built to maximize attack success rate (ASR). Yet ASR counts how often a target breaks, not in how many different ways. A red teamer can achieve a high ASR by reusing a few techniques that work, leaving the others untested and absent from the attacks used to harden models. We present AutoPI, an adaptive red teamer designed to find many different ways to break a target, not just to break it often. AutoPI uses PIAtlas, a taxonomy of prompt-injection techniques, both to guide its search and to measure coverage. For each scenario, an attacker LLM writes an injection for every sub-technique in the taxonomy and then reworks the ones that fail using feedback from an LLM judge. The method is black-box and training-free, it plugs into existing benchmarks in place of their fixed attack templates, and it stays up to date as the taxonomy evolves. We evaluate AutoPI on AgentDojo and IPI Arena against open-weight and proprietary target models. Alongside ASR, we report technique coverage, the fraction of taxonomy classes, techniques, and sub-techniques that yield at least one successful injection. On AgentDojo, AutoPI finds a successful injection against Qwen3.6-35B in all six classes of the taxonomy and for more than 90% of its techniques. Against Claude Haiku 5.5, the most robust target we evaluate, it achieves between 20% and 50% higher ASR and between 25% and 400% higher class-level coverage than recent state-of-the-art adaptive red teamers, while requiring no training. We release our code and the generated attacks to support reproducible, coverage-aware evaluation. 哦,原来用户给的正文是一整段?不对,那怎么分自然段落?等下,可能用户的原文是逻辑上的几个段落?不对,再看内容,逻辑上可以分成几个自然段: 1. 背景和现有问题:Prompt injection...harden models. 2. 提出AutoPI及方法:We present AutoPI...evolves. 3. 评估设置:We evaluate AutoPI...successful injection. 4. 评估结果:On AgentDojo...no training. 5. 开源说明:We release...evaluation. 对,这样分的话,每个用
包裹,然后重要的句子加strong,比如第一句的Prompt injection是核心风险,可以加strong?然后“We present AutoPI”也可以加strong?还有关键的性能数字那里?不对,要合理,不能乱加。 等下,再仔细看规则,重要句子可用,所以可以挑核心的: 比如第一句:Prompt injection is the dominant unsolved security risk for LLM agents, and the community now agrees that robustness must be measured against adaptive attackers rather than fixed benchmark injections. 然后后面的内容正常。 然后“We present AutoPI, an adaptive red teamer designed to find many different ways to break a target, not just to break it often.” 这个也加strong,因为是核心提出的对象。 还有结果里的“it achieves between 20% and 50% higher ASR and between 25% and 400% higher class-level coverage than recent state-of-the-art adaptive red teamers, while requiring no training” 这个也可以加,因为是关键结果。 然后,有没有需要列表的?原文里没有列表内容,所以不用ul。引用?没有。代码?没有。那