首页 > AI前沿 > CapLite: Prompt-Injection Defense for Small Language Model Agents without Model-Written Security Artifacts

CapLite: Prompt-Injection Defense for Small Language Model Agents without Model-Written Security Artifacts

arXiv自然语言 2026-11-13 01:07 24 阅读 查看原文
用户现在需要把给定的纯文本新闻正文排版成语义化HTML片段,按照规则来。首先看内容,首先整个是研究相关的内容,首先第一段是讲现有的防御方案和提出的问题对吧?然后接下来是实验结果,然后是提出的CapLite,然后是效果。 首先,先处理第一个自然段?不对,先拆分内容: 首先第一个部分,开头讲Enforcement defenses那段,是第一个

?然后接下来讲实验结果的那段是第二个

?然后讲CapLite的那段是第三个?不对,等下看原文的段落: 哦原文的段落是: 第一段:Enforcement defenses against prompt injection...We ask whether small language model (SLM) agents can be trusted to write these artifacts. 第二段:Across 15 open-weight models...and CaMeL's programs fail to run in up to 68% of episodes. 第三段:We present CapLite...that the user did not state. 第四段:No attack succeeds in 13,935 attacked AgentDojo episodes. 第五段:CapLite reaches utility under attack comparable to CaMeL's...of all enforcement defenses. 不对?等下再仔细读原文:哦不对,原文是连贯的?不对用户给的正文是: 第一句:Enforcement defenses against prompt injection, such as CaMeL, Progent, IPIGuard and DRIFT, ask the agent's own model to write a security artifact (a program, a privilege policy, a dependency graph or a constraint list) that code then enforces. 第二句:We ask whether small language model (SLM) agents can be trusted to write these artifacts. 然后第三句:Across 15 open-weight models, 13 of them between 1.7B and 12B parameters, they cannot: the policies, graphs and constraints that the agent's model writes admit attacks on 14 (Progent), 13 (IPIGuard) and 11 (DRIFT) of the models, and CaMeL's programs fail to run in up to 68% of episodes. 然后第四句:We present CapLite, which derives its policy from the tool schemas, commits the write calls before any tool result is read, and admits a write only if every argument that decides whom or what it affects is stated in the user's request. 第五句:This guarantees write integrity: injected text cannot add a write, and no write reaches a recipient, amount, object or destination that the user did not state. 第六句:No attack succeeds in 13,935 attacked AgentDojo episodes. 第七句:CapLite reaches utility under attack comparable to CaMeL's (29.0% vs. 27.2%) with fewer tokens than every other enforcement defense and than the undefended agent: 17.1k per episode, 18% of CaMeL's, and the highest utility per token of all enforcement defenses. 哦那拆分段落的话,首先: 第一个

:讲现有防御和提出的问题,就是前两句?不对,或者按内容逻辑分: 哦对,首先可以有

?不对,看内容,首先是不是可以分小节?不对,用户说长文用h2/h3划分,这里其实内容不长?不对,等下,首先,第一部分是背景与研究问题,然后是现有方案的评估结果,然后是提出的CapLite方案,然后是实验效果?不对,不过要保留所有信息,不得增删改。 等下,首先,先处理: 首先第一个

:Enforcement defenses against prompt injection, such as CaMeL, Progent, IPIGuard and DRIFT, ask the agent's own model to write a security artifact (a program, a privilege policy, a dependency graph or a constraint list) that code then enforces. We ask whether small language model (SLM) agents can be trusted to write these artifacts. 这部分是背景和问题,没问题。 然后第二个

:Across 15 open-weight models, 13 of them between 1.7B and 12B parameters, they cannot: the policies, graphs and constraints that the agent's model writes admit attacks on 14 (Progent), 13 (IPIGuard) and 11 (DRIFT) of the models, and CaMeL's programs fail to run in up to 68% of episodes. 这部分是评估结果,对。 然后接下来,第三个

:We present CapLite, which derives its policy from the tool schemas, commits the write calls before any tool result is read, and admits a write only if every argument that decides whom or what it affects is stated in the user's request. 这是方案介绍。 然后第四个

:This guarantees write integrity: injected text cannot add a write, and no write reaches a recipient, amount, object or destination that the user did not state. 这里重要内容可以加strong。 然后第五个

:No attack succeeds in 13,935 attacked AgentDojo episodes. 对。 然后第六个

:CapLite reaches utility under attack comparable to CaMeL's (29.0% vs. 27.2%) with fewer tokens than every other enforcement defense and than the undefended agent: 17.1k per episode, 18% of CaMeL's, and the highest utility per token of all enforcement defenses. 这里重要内容比如效果数据可以加strong? 不对,等下有没有需要用其他标签的?比如有没有列表?原文里没有列表,没有引用,没有代码,所以不用那些。然后有没有需要h2?哦原文是新闻正文,有没有小节?哦如果内容不长的话,是不是可以不用h2?不对,等下看内容,是不是可以分: 比如

现有防御方案的可靠性评估

?不对不对!用户说不得增删改内容!哦对!规则4是保留原文所有信息,不得增删改内容,所以不能加额外的小标题内容!哦对!我刚才差点犯错误,不能加原文没有的文字,所以h2/h3如果用的话,不能自己加文字,哦不对不对,原文里没有小标题,那是不是就不用h2/h3?对,因为原文没有小节标题,所以不需要加,否则就增加内容了。 哦对,规则4是保留原文所有信息,不得增删改内容,所以绝对不能加原文没有的文字,包括小标题。那所以全部用

,然后重要的句子加。 那现在整理: 首先第一个

:Enforcement defenses against prompt injection, such as CaMeL, Progent, IPIGuard and DRIFT, ask the agent's own model to write a security artifact (a program, a privilege policy, a dependency graph or a constraint list) that code then enforces. We ask whether small language model (SLM) agents can be trusted to write these artifacts. 这里没问题,有没有重要内容?比如研究问题可以加strong?比如We ask whether small language model (SLM) agents can be trusted to write these artifacts.?可以,因为是核心问题。 然后第二个

:Across 15 open-weight models, 13 of them between 1.7B and 12B parameters, they cannot: the policies, graphs and constraints that the agent's model writes admit attacks on 14 (Progent), 13 (IPIGuard) and 11 (DRIFT) of the models, and CaMeL's programs fail to run in up to 68% of episodes. 这里的评估结果可以加strong吗?比如Across 15 open-weight models, 13 of them between 1.7B and 12B parameters, they cannot: 后面内容不变。 然后第三个

:We present CapLite, which derives its policy from the tool schemas, commits the write calls before any tool result is read, and admits a write only if every argument that decides whom or what it affects is stated in the user's request. 这里We present CapLite 加strong? 然后第四个

:This guarantees write integrity: injected text cannot add a write, and no write reaches a recipient, amount, object or destination that the user did not state. 这里的核心保证加strong。 然后第五个

:No attack succeeds in 13,935 attacked AgentDojo episodes. 这个实验结果加strong。 然后第六个

:CapLite reaches utility under attack comparable to CaMeL's (29.0% vs. 27.2%) with fewer tokens than every other enforcement defense and than the undefended agent: 17.1k per episode, 18% of CaMeL's