新框架从能力增益、影响范围与部署难度为 AI 越狱评级

最近几个月,关于 AI 模型摆脱自身防护机制的重磅新闻占据了头条。如今,Frontier Model Forum(FMF)发布了一套新的越狱评级框架(Jailbreak Rating Framework),旨在追踪某个特定 AI 模型绕过防护机制究竟有多容易。该框架还评估了若某个 AI 模型在互联网上造成网络破坏,其附带损害可能达到的规模。
该项目于 2023 年启动,得到四家创始公司的支持:Anthropic、Google、Microsoft 和 OpenAI。此次发布的越狱评级框架大致相当于知名的 CVSS 安全评级系统,但专门用于在多个维度上、并覆盖多家前沿 LLM 提供商,对 LLM 越狱的类型和严重程度进行分类。
许多 AI 公司和评估机构已经有自己的评级系统,但 FMF 希望其量表能成为跨厂商标准,供开发补丁的团队依赖。
FMF 的思路是为越狱建立一种分类方法,用以衡量模型所能执行的网络战操作是否超出目前正规市场和黑市上已有工具的水平。该提案的 1.0 版仍是一项进行中的工作,该组织在 12 月 21 日之前持续征集意见和建议,并指出在此期间不宜依赖其系统。
目前,该框架有四个评分组成部分,分为两大类别:严重性,衡量某种越狱技术的严重程度;可能性,即发现并将其武器化的难易程度。
综合最终评级会把这两个类别所得的分数相加,得出一个从低到危急的最终评级。科学家指出,能力增益是最重要的维度,它同时构成其余各项分数的下限和上限——如果一种越狱无法带来超越通过其他手段已能获得的结果,那么在评分体系的考量范围内,它就没有意义。
各项单独评级分别涉及能力增益、能力范围、部署难易程度和发现难易程度。它们都按 0 到 3 分打分,不过 FMF 强调,评分为 0 并不意味着某个越狱无害或应当被忽视——而是意味着该缺陷应由标准做法来处理,比如平台的政策和标准网络安全评级。
一个重要特点是,FMF 的基线会随着模型演进隐式地变化,这意味着 2026 年某个越狱评为 0 分,其含义很可能与 2028 年的 0 分不同,或者说并不优于其他替代工具。能力增益评级为 0 意味着,所分析的越狱相比现有模型或工具一上手就能做到的事情,没有任何改进。
例如,越狱模型对一个网站发起 DDoS 攻击对攻击者来说帮助不大,因为已有不少地下黑市网站提供此类服务——甚至还有其他防护更少的 LLM。这并不意味着该越狱不好用,只是说它并不比替代方案更好。不过,如果攻击者能利用该越狱实施比市面上能买到的更强的 DDoS,那么评级就可能更高。
在能力增益量表的另一端,三级意味着被越狱的模型能够以与无防护机制运行的模型相近的能力实施网络攻击。
能力范围的衡量更容易理解:零级意味着该越狱手段只对某一个具体操作或目标有效,无法针对其他用途进行调整,换言之,只是一招鲜。当一种越狱手段无需专门重新设计,就能让模型在至少两个完全不相干的网络战类别中开展工作——比如既能让模型自主发现漏洞,又能单独让它编写恶意软件——它就会被评为三级。
在易用性方面,FMF 考量的是所需的人工介入程度,也就是该越狱手段所允许的自动化程度。零级意味着该越狱必须由具备高度专业知识的人来操控,还要提供额外指令才能生效;而三级则意味着新手几乎不费力气就能使用,并能自动完成一项网络战任务。
最后但同样重要的一点是,易发现性衡量的是找到该越狱手段的难度。若为零级,攻击者必须花费大量资金购买推理 Token,需要更高权限(例如授予白帽黑客的特权账户),或需要 AI 服务器集群这类专门基础设施。评为三级则仅仅意味着任何人都能查到这一技术并加以使用。
如果这一切听起来似曾相识,你可能会想到 Anthropic 在 2026 年 7 月提出的防护机制失效分类器——此前一起事件曾导致 Claude Fable 5 暂时不可用。与 FMF 的做法十分相似,Anthropic 的系统同样会采用动态基线,并且只把那些相较于现有网络战能力有所提升的越狱纳入考量。
那份文件与 FMF 的提案惊人地相似。Anthropic 是 FMF 的创始成员,因此可能在一定程度上参与了这个新标准的制定——不过我们无法证实这一点。
译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。
查看英文原文(New framework grades AI jailbreaks across capability gains, scope, and how easy they are to deploy)
In recent months, high-profile headlines about AI models letting loose from their safeguards dominated headlines. Now, the Frontier Model Forum (FMF) has released a new Jailbreak Rating Framework that aims to track how easy it might be for a given AI model to get around safeguards. The framework also estimates the scale of the collateral damage, should an AI model wreak cyber-havoc across the internet.
The project started in 2023 with the support of four founding companies: Anthropic, Google, Microsoft, and OpenAI. The published Jailbreak Rating Framework is a rough analog to the well-known CVSS security rating system, but specifically targeted at classifying the type and severity of LLM jailbreaks across various axes and multiple frontier LLM providers.
Many AI companies and evaluation entities already have their own rating systems, but the FMF intends for its scale to be a cross-vendor standard that teams developing patches can rely on.
The FMF's approach is to have a classification method for jailbreaks that can make a model perform cyber-warfare operations beyond what's already out, across both the regular and black market. Version 1.0 of the proposal is a work-in-progress, and the group is taking comments and suggestions on it until December 21, noting that relying on its system in the meantime is inadvisable.
Right now, the framework has four scoring components, split across two main categories: severity, measuring how serious a jailbreak technique is, and likelihood, or how easy it is to find and weaponize it.
A combined final rating adds the scores obtained from the two categories into a final low-to-critical rating. The scientists note that the capability gain is the most important dimension and acts as both a floor and a ceiling for the remaining scores — a jailbreak that doesn't offer results beyond what can already be obtained via other means is meaningless within the purview of the grading system.
The individual ratings pertain to capability gains, capability scope, ease of deployment, and ease of discovery. They're all graded from zero to three, though the FMF stresses that a rating of zero doesn't mean a jailbreak is harmless or ought to be ignored — rather, it means the flaw should be addressed by standard practices like the platform's policies and standard cybersecurity ratings.
An important feature is that the FMF's baseline implicitly changes as models evolve, meaning that a grade of zero for a jailbreak in 2026 likely isn't going to mean the same thing as it would in 2028, or isn't better than alternative tools. A capability-gain rating of zero means that the jailbreak under analysis doesn't offer any improvement beyond what can already be done with existing models or tooling right off the bat.
For example, jailbreaking a model to perform a DDoS on a site isn't all that helpful to an attacker, since there are a number of black-market sites for that — or even other LLMs with fewer safeguards. This doesn't mean the jailbreak isn't handy, only that it's no better than the alternatives. However, the rating can be higher if the attacker can leverage the jailbreak to get a better DDoS than they could otherwise shop for.
On the other end of the capability-gain scale, three means that a jailbroken model can perform cyber-attacks with capabilities similar to those of a model operating without safeguards.
The capability scope measurement is easier to interpret: zero means that the jailbreak only works for one specific action or target and can't be tweaked for other applications, or in other words, is a one-trick pony. A jailbreak gets a rating of three when it can, without specific redesign, make a model perform work across at least two entirely unrelated categories of cyber-warfare — like getting the model to both autonomously discover vulnerabilities or, separately, write malware.
For ease of deployment, the FMF considers the amount of human interaction, or the degree of automation that the jailbreak allows. A zero means that the jailbreak must be steered by humans with a high degree of knowledge, providing additional instructions to be effective, while a three means that, with little effort, it can be used by a novice and results in an automated cyber-warfare task.
Last but by no means least, ease-of-discovery is how hard it is to find the jailbreak. At a zero, the attacker must spend copious amounts of money on inference tokens, need elevated access (like a privileged account given to a white-hat hacker), or specialized infrastructure like a legion of AI servers. A rating of three simply means that anyone can look up the technique and use it.
If all of this sounds familiar, you may be thinking of Anthropic's safeguard-breakage classifiers the company proposed in July 2026, after an incident that led to Claude Fable 5 being temporarily unavailable. Much like the FMF's effort, Anthropic's system would also use a moving baseline and only consider jailbreaks that improve on cyber-warfare capabilities already available.
That document is strikingly similar to FMF's proposal. Anthropic is an FMF founding member, and as such may have had a hand or two in developing the new standard — though we have no confirmation either way.
正文保留英文原文,标题与摘要由机器翻译。
原文发布于 Tom's Hardware(New framework grades AI jailbreaks across capability gains, scope, and how easy they are to deploy),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。