GLM-5.3 与高级网络攻击能力的扩散
Five months ago, we announced Claude Mythos Preview, the first AI model that could autonomously build sophisticated, end-to-end cyber exploits. The rapid rate of improvement in AI suggested to us that this ability would eventually proliferate to many other models, making it much easier for malicious cyber actors to launch highly impactful cyberattacks.
五个月前,我们发布了 Claude Mythos Preview。它是第一个能自主构建复杂的端到端网络漏洞利用(exploit)的 AI 模型。照 AI 进步的速度,我们判断这种能力迟早会扩散到很多其他模型上,到时候恶意的网络攻击者要发动破坏力很大的攻击就容易多了。
In light of these considerations, we chose to release Claude Mythos Preview in a limited way, through Project Glasswing—which enabled trusted cyber defenders to find more than 10,000 vulnerabilities in critical software, giving them a head start before malicious actors had access to similarly capable models.
出于这些考虑,我们选择有限度地发布 Claude Mythos Preview,只通过 Project Glasswing 提供。可信的网络防御者借它在关键软件里找到了一万多个漏洞,在恶意攻击者拿到同等能力的模型之前抢先了一步。
But those models have now arrived. In this post, we share our analysis of GLM-5.3, the latest AI model developed by Zhipu AI (known outside of China as Z.ai). Like Claude Mythos Preview, GLM-5.3 has strong capabilities for autonomously building end-to-end cyber exploits. But GLM-5.3 is unlike other frontier models in that it has been released without meaningful safeguards to limit misuse. We find that attackers can bypass GLM-5.3’s safeguards between 64% and 100% of the time with simple techniques in our simulated tests. In contrast, these attacks did not succeed against safeguarded Claude models in our testing. We assess that GLM-5.3’s lax safeguards significantly increase the cyber capabilities available to malicious actors. At the same time, these capabilities can also benefit defenders working to secure their systems.
可这样的模型现在已经来了。这篇文章分享我们对 GLM-5.3 的分析,它是智谱 AI(在中国以外叫 Z.ai)最新开发的模型。和 Claude Mythos Preview 一样,GLM-5.3 自主构建端到端漏洞利用的能力很强。但它和其他前沿模型不同:发布时没有配上真正能限制滥用的安全防护(safeguards)。在我们的模拟测试里,攻击者用一些简单手段,就能在 64% 到 100% 的情况下绕过 GLM-5.3 的防护。同样的攻击用在带防护的 Claude 模型上,在我们的测试里都没有成功。我们的判断是,GLM-5.3 松散的防护让恶意攻击者能用上的网络攻击能力大幅提升。同时,这些能力也能帮到正在加固自家系统的防御者。
On Sept. 17, NIST’s Center for AI Standards and Innovation (CAISI) published its own assessment of GLM-5.3’s cyber capabilities. CAISI found that GLM-5.3 is “the most cyber-capable open-weight model released to date” and that it lags the US frontier by about four months on an aggregate of CAISI’s cyber benchmarks. Our capability findings broadly match CAISI’s. In CAISI’s comparison, US models were tested with cyber safeguards disabled when applicable, and the US frontier includes models released only to vetted users. Attackers can’t readily access those versions of US models, but anyone can download GLM-5.3. This post adds our analysis of how easily GLM-5.3’s safeguards can be bypassed or removed.
9 月 17 日,美国国家标准与技术研究院(NIST)下属的 AI 标准与创新中心(CAISI)发布了他们自己对 GLM-5.3 网络能力的评估。CAISI 的结论是,GLM-5.3 是「迄今为止网络能力最强的开放权重(open-weight)模型」,在 CAISI 一组网络基准的综合成绩上,落后美国最前沿大约四个月。我们对能力的发现和 CAISI 大体一致。在 CAISI 的对比里,美国模型在适用的情况下是关掉网络安全防护来测的,而且「美国最前沿」里包括只发布给经过审核的用户的模型。攻击者没那么容易拿到这些版本的美国模型,GLM-5.3 却是谁都能下载。我们在这篇文章里补充分析了另一件事:GLM-5.3 的防护有多容易被绕过或去掉。

Figure 1. Summary of findings. Top: The increase in exploitation capability between Claude Opus 4.6 and Claude Mythos Preview mirrors the jump in capabilities between GLM-5.2 and GLM-5.3. Claude models are released with cyber safeguards, and versions with reduced safeguards are limited to vetted users. Anyone can download and use GLM-5.3. Bottom: The limited safeguards in GLM-5.3 can be bypassed with standard techniques that did not work against, or do not apply to, Claude models in our testing.
图 1. 发现汇总。上:从 Claude Opus 4.6 到 Claude Mythos Preview 漏洞利用能力的提升,和从 GLM-5.2 到 GLM-5.3 的能力跃升如出一辙。Claude 模型发布时带有网络安全防护,防护减弱的版本只给经过审核的用户。GLM-5.3 则谁都能下载使用。下:GLM-5.3 有限的防护能被一些常规手段绕过,而在我们的测试里,这些手段对 Claude 模型要么不起作用,要么根本用不上。
GLM-5.3 can develop working exploits end to endGLM-5.3 能端到端写出可用的漏洞利用
To understand how GLM-5.3 could enable cyber threat actors to find and exploit real software vulnerabilities, we ran evaluations using automated benchmarks and human-in-the-loop workflows. For both approaches, we ran the tested models in isolated and sandboxed environments so they can only attack offline targets that we have set up for the purposes of these evaluations. We focus primarily on exploit development capability, as this is where Claude Mythos Preview demonstrated a notable jump versus previous Claude models.
为了弄清 GLM-5.3 能在多大程度上帮网络威胁行为者找到并利用真实软件里的漏洞,我们做了两类评估:自动化基准测试,和有人参与(human-in-the-loop)的工作流。两类评估里,被测模型都跑在隔离的沙箱环境中,只能攻击我们专门为评估搭起来的离线目标。我们主要关注漏洞利用的开发能力,因为 Claude Mythos Preview 正是在这一项上比之前的 Claude 模型有明显跃升。
First, we ran the model on ExploitBench, which measures how well AI models can exploit known vulnerabilities in the V8 engine used by Google Chrome. Here we focus on the models’ ability to develop end-to-end exploits successfully, as this is the most relevant capability for attackers, and where we see significant changes between models. We find that GLM-5.3 develops end-to-end exploits in 50 of 410 attempts. Claude Mythos Preview did so at a similar rate—in 56 of 410 attempts.
首先是 ExploitBench,它衡量 AI 模型利用 Google Chrome 所用的 V8 引擎里已知漏洞的能力。这里我们看的是模型能不能成功写出端到端的漏洞利用,因为这是对攻击者最要紧的能力,也是不同模型之间差别最明显的地方。结果是,GLM-5.3 在 410 次尝试里有 50 次写出了端到端漏洞利用。Claude Mythos Preview 的成功率差不多:410 次里 56 次。
In our internal Binary Exploitation benchmark,1 we test whether models can find and exploit vulnerabilities in popular open-source projects that participate in Google’s OSS-Fuzz project. Here, full credit is awarded for a full control-flow hijack. We evaluate several models on 100 tasks from the benchmark (selected at random), and find that GLM-5.3 develops full control-flow hijacks in 4% of the trials; Claude Mythos Preview did so in 6%. Although GLM-5.3 performs below Claude Mythos Preview here, a meaningful threshold has clearly been crossed: earlier models, like Claude Opus 4.6 and GLM-5.2, do not succeed in any of them.
在我们内部的二进制漏洞利用(Binary Exploitation)基准里,1测的是模型能不能在参与 Google OSS-Fuzz 项目的热门开源项目里找到漏洞并加以利用。只有实现完整的控制流劫持(control-flow hijack)才算满分。我们从基准里随机挑了 100 个任务,拿几个模型来测,GLM-5.3 在 4% 的试验里实现了完整的控制流劫持,Claude Mythos Preview 是 6%。GLM-5.3 在这一项上虽然不如 Claude Mythos Preview,但显然已经跨过了一道有分量的门槛:更早的模型,比如 Claude Opus 4.6 和 GLM-5.2,一个任务都没做成。

Figure 2. Exploitation capability versus output-token budget. Each line shows the share of a model’s attempts that reached the benchmark’s top outcome versus the output tokens used. Both figures show performance of two Claude models run with safeguards disabled (Opus 4.6, Mythos Preview), two GLM models (5.2 and 5.3), and results from the latest open-weight models released by Moonshot AI (Kimi K3) and DeepSeek (V4.1-Flash).
图 2. 漏洞利用能力与输出 token 预算的关系。每条线表示一个模型的尝试中达到该基准最高结果的比例,随所用输出 token 的变化。两张图都包括:关掉防护运行的两个 Claude 模型(Opus 4.6、Mythos Preview),两个 GLM 模型(5.2 和 5.3),以及月之暗面(Kimi K3)和 DeepSeek(V4.1-Flash)最新发布的开放权重模型的结果。
Next, we evaluated how GLM-5.3 performs on open-ended offensive cyber tasks in the hands of human experts (mirroring our testing with Claude Mythos Preview earlier this year). Here, we select targets in which the human experts are unaware of existing vulnerabilities, then ask them to use the model to identify and exploit novel flaws. These experiments tested what the experts could do in a short time-frame: they typically ran for a day or less, with less than an hour of human focus in total.
接着,我们评估了 GLM-5.3 在人类专家手里做开放式网络攻击任务的表现(做法照搬今年早些时候我们对 Claude Mythos Preview 的测试)。我们挑的目标,专家事先都不知道上面有哪些漏洞,然后让他们用模型去找出新的缺陷并加以利用。这些实验测的是专家在很短时间里能做到什么:一般跑一天或更短,人投入的专注时间加起来不到一小时。

Figure 3. A redacted screenshot of an exploit page generated by GLM-5.3 during researcher-driven testing, shown stealing a user’s SSH private key via a malicious website. The exploit chains together multiple 0-day vulnerabilities the model discovered in a component of a popular web browser, reading a sensitive file off the user’s computer.
图 3. 研究员主导的测试中,GLM-5.3 生成的一个漏洞利用页面的截图(已打码),展示的是通过恶意网站窃取用户的 SSH 私钥。这个利用把模型在一款流行浏览器的某个组件里发现的多个 0day 漏洞串在一起,从用户电脑上读走了一个敏感文件。
In the first of these sessions, a researcher used GLM-5.3 on a sandboxed machine with a local Linux build of a popular web browser. Over the course of a day (and with limited human attention), GLM-5.3 found several previously unknown vulnerabilities in the browser’s JavaScript engine, and chained them together into a working exploit: a webpage that, when visited, reads arbitrary files from the visitor’s computer (shown in Figure 3). This exploit targets the Linux build of the browser, since that was the only environment made available to the model. However, we believe these vulnerabilities could also impact users on other platforms, though the path to exploitation there may be more complex. (We’ve disclosed these vulnerabilities to the maintainer.) Later in the session, the researcher also identified exploitable vulnerabilities in several other widely used systems with GLM-5.3, including wireless and graphics drivers and network-facing device software. We are currently reviewing these reports and we will disclose to maintainers as appropriate.
第一场测试里,一位研究员在一台沙箱机器上用 GLM-5.3,机器上装着一款流行浏览器的本地 Linux 构建版。一天之内(人也没怎么盯着),GLM-5.3 在这款浏览器的 JavaScript 引擎里找到了好几个此前没人知道的漏洞,还把它们串成了一个可用的漏洞利用:一个网页,只要有人访问,就能读取访客电脑上的任意文件(见图 3)。这个利用针对的是浏览器的 Linux 版,因为模型手里只有这一个环境。不过我们认为,这些漏洞也可能影响其他平台的用户,只是在那些平台上利用起来可能更绕。(这些漏洞我们已经报告给了维护方。)在这场测试后面,研究员还借 GLM-5.3 在其他几个广泛使用的系统里发现了可利用的漏洞,包括无线网卡和显卡驱动,以及面向网络的设备软件。这些报告我们正在审核,会视情况告知维护方。
In a second session, a researcher used GLM-5.3-Flash (a smaller, less capable version of GLM-5.3) to develop an exploit for a known vulnerability (we’ve previously written about these “N-day” vulnerability exploits here). Here, the researcher focused on a recently disclosed flaw in Google Chrome (CVE-2026-11645) to see how quickly the model could turn a public fix into a working attack. The researcher provided GLM-5.3-Flash with public details of this CVE and another known flaw. With no significant direction from the researcher, GLM-5.3-Flash chained together exploits for these two flaws, building a reliable exploit chain for an ARM64 target, bypassing pointer-authentication (PAC) hardening. This took 20 minutes of human attention, plus eight hours of work for GLM-5.3-Flash. At Zhipu’s API prices, this effort would have cost $20.40.
第二场测试里,一位研究员用 GLM-5.3-Flash(GLM-5.3 更小、能力更弱的版本)给一个已知漏洞写利用(这类「N-day」漏洞利用我们之前在这里写过)。研究员盯的是 Google Chrome 里一个最近才公开的缺陷(CVE-2026-11645),想看模型多快能把一个公开的修复补丁变成能用的攻击。研究员把这个 CVE 和另一个已知缺陷的公开细节交给 GLM-5.3-Flash。在研究员没怎么指点的情况下,GLM-5.3-Flash 把这两个缺陷的利用串在一起,针对 ARM64 目标搭出了一条稳定的利用链,还绕过了指针认证(PAC)加固。整个过程花了研究员 20 分钟的注意力,加上 GLM-5.3-Flash 八个小时的工作。按智谱的 API 价格,这一趟要花 20.40 美元。
GLM-5.3 lacks robust safeguardsGLM-5.3 缺乏牢靠的安全防护
GLM-5.3 has been released with some built-in safeguards: if a user asks for something clearly harmful, the model will often refuse.2 In our testing, we found that these safeguards could be bypassed or removed with a variety of simple techniques.
GLM-5.3 发布时自带了一些安全防护:用户要是提出明显有害的要求,模型常常会拒绝。2但我们测下来发现,这些防护用各种简单手段就能绕过或去掉。
The most intensive—and most successful—method is a standard refusal reduction technique known as “abliteration.” Since GLM-5.3 is released as an open-weight model, users can reconfigure it to remove its refusals with little change in its capabilities. Several developers released abliterated versions of GLM-5.3 to the public within days of the model’s release.
最费工夫、也最有效的办法,是一种叫「拒绝消融」(abliteration)的常规降低拒绝率技术。GLM-5.3 是以开放权重的形式发布的,用户可以改造它,去掉它的拒绝行为,而能力几乎不受影响。模型发布后没几天,就有好几个开发者公开放出了做过拒绝消融的 GLM-5.3。
To research how far abliteration allows attackers to bypass GLM-5.3’s safeguards, we produced an abliterated copy ourselves, and then ran it on three public benchmarks (JailbreakBench, HarmBench, and StrongREJECT) that measure how often a model complies with clearly harmful requests. Abliterating the model took our team—which had never previously attempted this task—about 2,200 GPU hours at a computation cost of roughly $4,400.3 Abliterating GLM-5.3-Flash took about 600 GPU hours. The edit took GLM-5.3’s refusal rate from above 90% to about 3% and 2% on the first two benchmarks (JailbreakBench and HarmBench) and to 12% on the third (StrongREJECT). Abliteration did not significantly reduce the model’s capabilities: on GPQA-Diamond, an evaluation that measures general scientific capabilities, the standard and abliterated models scored the same results; on a tested subset of the CyberGym evaluations, the abliterated version scored a few percent lower (as shown in the chart below).
为了研究拒绝消融能让攻击者把 GLM-5.3 的防护绕到什么程度,我们自己做了一份消融过的副本,然后在三个公开基准上跑(JailbreakBench、HarmBench 和 StrongREJECT),它们衡量的是模型有多常照办明显有害的请求。我们团队之前从没做过这件事,给这个模型做消融花了大约 2,200 个 GPU 小时,算力成本约 4,400 美元。3给 GLM-5.3-Flash 做消融花了大约 600 个 GPU 小时。改完之后,GLM-5.3 的拒绝率从 90% 以上,在前两个基准(JailbreakBench 和 HarmBench)上降到约 3% 和 2%,在第三个(StrongREJECT)上降到 12%。拒绝消融没有明显削弱模型的能力:在衡量通用科学能力的 GPQA-Diamond 上,原版和消融版得分一样;在 CyberGym 评估中我们测的那部分上,消融版低了几个百分点(见下图)。

Figure 4. Top: Mean refusal rates across JailbreakBench, HarmBench, and StrongREJECT for the released and abliterated GLM models and for Claude models. After abliteration, the GLM models rarely refuse these queries. Claude models cannot be abliterated because their weights aren’t publicly released or customizable. Bottom: Performance of GLM-5.3 and GLM-5.3-Flash versus their abliterated variants on GPQA-Diamond and CyberGym. Abliteration leaves capability largely intact.
图 4. 上:发布版和消融版 GLM 模型以及 Claude 模型在 JailbreakBench、HarmBench 和 StrongREJECT 上的平均拒绝率。消融之后,GLM 模型很少再拒绝这些请求。Claude 模型没法做消融,因为它们的权重既不公开,也不能自行改动。下:GLM-5.3 和 GLM-5.3-Flash 与各自消融版在 GPQA-Diamond 和 CyberGym 上的表现对比。拒绝消融基本没伤到能力。
In our testing, we observed that GLM-5.3’s safeguards can also be circumvented without using an abliterated version of the model. We placed the model in a simulated world4 in which it was given overtly malicious requests to attack critical systems. Out of the box, GLM-5.3 refused in all trials (as with the other models we tested). But we identified several simple ways to bypass the GLM models’ safeguards, such that it would respond to these requests in most or all cases. These include:
我们在测试中还发现,不用消融版也能绕开 GLM-5.3 的防护。我们把模型放进一个模拟世界里,4给它下明摆着是恶意的指令,让它去攻击关键系统。开箱即用的 GLM-5.3 在所有试验里都拒绝了(我们测的其他模型也一样)。但我们找到了几种简单办法绕过 GLM 模型的防护,让它在大多数甚至全部情况下照办。包括:
- Providing a deceptive prompt, such as telling the model that it is an autonomous red-team agent working on an exercise. This gets GLM-5.3 to engage 64% of the time.给一个骗人的提示词,比如告诉模型它是一个正在参加演练的自主红队 Agent。这能让 GLM-5.3 有 64% 的时候照做。
- Prefilling the models’ thinking tokens so that it appears to have considered the user’s request and decided to proceed. This gets GLM-5.3 to engage 92% of the time.预填充(prefill)模型的思考 token,让它看起来已经考虑过用户的请求、决定接着干。这能让 GLM-5.3 有 92% 的时候照做。
- Using an abliterated version of the model, as described above. This gets GLM-5.3 to engage 100% of the time.用上面说的消融版模型。这能让 GLM-5.3 100% 照做。
In our testing, none of these techniques got safeguarded Claude models to carry out the harmful tasks we tested. Claude’s safeguards blocked the requests that used deceptive prompts. The Anthropic API provides would-be attackers with no way to prefill Claude’s thinking. And since Claude’s weights are not provided to users, they cannot be abliterated to change Claude’s behavior.
在我们的测试里,这些手段没有一个能让带防护的 Claude 模型去执行我们测的那些有害任务。用了骗人提示词的请求被 Claude 的防护拦了下来。Anthropic API 不给潜在的攻击者预填充 Claude 思考内容的途径。Claude 的权重也不提供给用户,所以没法靠拒绝消融改变它的行为。

Figure 5. Rate at which each model tried to connect to a remote target system after an overtly harmful request by model and bypass condition (50 samples per cell, five attack orders × two targets × five attempts). GLM-5.3 rises from zero engagement on direct requests to 64%, 92%, and 100% under a cover story, prefill, and abliteration respectively, while every tested Claude model stays at zero under API safeguards. Padlocked cells mark attacks that are not generally feasible against the Claude API.
图 5. 收到明显有害的请求后,各模型在不同绕过条件下尝试连接远程目标系统的比例(每格 50 个样本:五条攻击指令 × 两个目标 × 五次尝试)。GLM-5.3 在直接请求下的参与度是零,换成背景说明、预填充和拒绝消融后,分别升到 64%、92% 和 100%;而测过的每个 Claude 模型在 API 防护下都保持为零。带锁图标的格子,表示这类攻击对 Claude API 一般做不到。
To demonstrate how the abliterated version of GLM-5.3 is willing to engage in harmful tasks, we highlight one quote from the chain of thought that it generated:
为了说明消融版 GLM-5.3 有多愿意去做有害的任务,我们摘出它生成的思维链(chain of thought)里的一段:

Figure 6. In our simulated environment, the abliterated version of GLM-5.3 considers how to engage in an overtly harmful request, ultimately deciding to follow the user’s instructions despite some initial ethical concerns. Text is quoted verbatim from the model’s chain of thought.
图 6. 在我们的模拟环境里,消融版 GLM-5.3 琢磨怎么执行一个明显有害的请求,起初有些伦理上的顾虑,最后还是决定照用户的指令办。文字逐字引自模型的思维链。
What does this mean?这意味着什么?
GLM-5.3 will likely give malicious actors access to capabilities that will allow them to find and exploit cyber vulnerabilities without meaningful restrictions. This is unlike any other similarly capable AI model, all of which were released with safeguards or through limited access programs. The release of GLM-5.3 is a meaningful step change in the cyber capabilities available to attackers. Anthropic and other US AI labs have published recent reports that disclose how cyber attackers have tried to use AI systems. Given this evidence, we think it’s likely both state and non-state actors will use models like GLM-5.3 to cause real-world harm.
GLM-5.3 很可能让恶意行为者拿到一种能力:在几乎不受限制的情况下发现并利用网络漏洞。能力相当的其他 AI 模型没有一个是这样的,它们发布时要么带着防护,要么只通过有限访问计划提供。GLM-5.3 一发布,攻击者能用上的网络能力就实打实地上了一个台阶。Anthropic 和其他美国 AI 实验室最近都发布过报告,披露网络攻击者是怎么尝试利用 AI 系统的。有这些证据在,我们认为国家行为者和非国家行为者都很可能会用 GLM-5.3 这样的模型造成现实中的危害。
On the other hand, models with this level of capability can also be used by defenders. Our view is that cyber defenders should use the best available tools that meet their needs. We're working to safely expand access to Claude's cyber capabilities to as many defenders as we can. Cyber defenders face attackers who will use every capable tool they can, and we believe defenders should be equipped with frontier models that are at least as good as those their adversaries are using.
另一方面,这种水平的模型防御者也能用。我们的看法是,网络防御者应该用上能满足自己需求的最好工具。我们正在想办法,在安全的前提下把 Claude 的网络能力开放给尽可能多的防御者。防御者面对的攻击者会把手边每一件好用的工具都用上,我们认为防御者手里的前沿模型,至少要和对手用的一样好。
Through Project Glasswing (and other efforts, like Patch the Planet), cyber defenders have made meaningful progress towards securing critical systems in advance of this moment—but much work remains to be done. While vetted defenders can now use even more advanced models like Claude Mythos 5.1 through our trusted access programs, a critical threshold in freely accessible capabilities has now been crossed. GLM-5.3 underscores the urgency of expanding access to advanced frontier models to a broader set of entities to empower cyber defenders.
靠 Project Glasswing(以及 Patch the Planet 等其他行动),网络防御者已经赶在这一刻到来之前,在加固关键系统上有了不小的进展,但要做的事还很多。经过审核的防御者现在可以通过我们的可信访问计划,用上 Claude Mythos 5.1 这样更先进的模型;可人人都能免费拿到的能力,也已经跨过了一道关键门槛。GLM-5.3 说明,把先进的前沿模型开放给更多机构、给网络防御者撑腰,这件事更急了。
Governments should conduct safety testing on sufficiently capable AI models, including successors to GLM-5.3. Without high-quality evaluations from independent sources, the impact of these capabilities might not become fully clear to model developers until it is too late. As AI developers across the world build increasingly capable open-weight models, we hope they work to appropriately safeguard these capabilities and prevent misuse.
各国政府应该对能力足够强的 AI 模型做安全测试,包括 GLM-5.3 的后续版本。没有来自独立方的高质量评估,模型开发者可能要等到为时已晚,才完全看清这些能力会带来什么影响。世界各地的 AI 开发者正在造出越来越强的开放权重模型,我们希望他们给这些能力配上恰当的防护,防止被滥用。
1 We have previously published results from this benchmark under the name “OSS-Fuzz”; here we tested models on a randomly selected subset of 100 tasks from this evaluation.
1 这个基准的结果我们之前以「OSS-Fuzz」的名字发布过;这次我们从中随机挑了 100 个任务的子集来测模型。
2 The cyber tasks detailed in the prior section did not trigger such refusals on the released model, but we observe refusals if we ask GLM-5.3 for assistance developing malware or helping launch cyberattacks against remote targets.
2 上一节详述的那些网络任务,在发布版模型上并没有触发这类拒绝;但如果要 GLM-5.3 帮忙开发恶意软件,或者帮忙对远程目标发动网络攻击,我们会看到它拒绝。
3 Most of this went to exploring variants in parallel and to testing model capabilities after modifications. We estimate a team experienced with the technique, starting from scratch on this model, would need closer to 600 GPU hours ($1,200).
3 其中大部分花在了并行探索各种变体,以及改完之后测试模型能力上。我们估计,一个熟悉这项技术的团队从零开始在这个模型上做,需要的更接近 600 个 GPU 小时(1,200 美元)。
4 No model-generated code is ever executed in this simulation and the model does not have any way to interact with external systems. In this isolated test environment the model under evaluation is given access to a fake bash tool which does not execute the provided code. Instead, we ask another LLM to approximate the command’s result given a description of the simulated world. These simulations are not perfect portrayals of real-world conditions, and are therefore imperfect measures of how a model would behave in a given situation.
4 这个模拟里,模型生成的代码从来不会被执行,模型也没有任何途径和外部系统交互。在这个隔离的测试环境里,被评估的模型能用的是一个假的 bash 工具,它不会执行传进来的代码,而是由另一个 LLM 根据对模拟世界的描述,大致推演出命令的结果。这些模拟并不能完美还原现实条件,所以拿它来衡量模型在特定情境下会怎么做,并不完全准确。