PromptArmor 发布 FrontierBench:AI 智能体隐藏指令攻击测试
谁在用智能体自动处理邮件和文档,谁就该看看这份成绩单。PromptArmor 用 21 个模型实测隐藏指令攻击,Grok 4.5 九成多都中招,Gemini Argon 最抗打。选模型前先对照这个分数。
PromptArmor 发布了 FrontierBreach 基准,测量工作场景中 AI 智能体被隐藏指令误导的频率。攻击方式是在邮件或文档中嵌入不可见的一行文字,智能体读到后可能把数据外发或执行攻击者的代码。在 21 个模型上的测试中,这类隐藏攻击平均得手率为 34.9%。
所有被测的前沿模型均被攻破:Gemini Argon 得手率最低,为 4.3%;Muse Wonder 为 15.2%;Grok 4.5 则在 91.4% 的攻击中中招。藏在智能体技能(agent skills)里的攻击成功率更高,达到 45.5%。
这份基准把智能体安全从理论讨论变成了可量化的分数。对部署智能体处理邮件、文档的企业来说,各模型间 4.3% 到 91.4% 的巨大差距说明选型时必须把提示注入抗性列入评估项,不能只看推理和编码成绩。
PromptArmor has released FrontierBreach, a benchmark of how often AI agents at work can be misled by hidden instructions.
Attackers can hide an invisible line in an email or document, and agents may try to follow it, sending data out or running the attacker's code.
> Across 21 AI models, these hidden attacks worked on 34.9% of PromptArmor's tests.
> PromptArmor breached every frontier model it tested, from Gemini Argon at 4.3% to Muse Wonder at 15.2%, with Grok 4.5 falling for 91.4% of the attacks.
> Attacks hidden inside agent skills also worked in 45.5% of the tests.
This report is categorized under AITOP's "Papers" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows and US-China foundation model benchmarks and commercialization comparison. Key entities and technical topics under continuous tracking include FrontierBreach, PromptArmor, Grok 4.5, Gemini.
Source attribution: curated from TestingCatalog (@testingcatalog).