跳到主内容
@wquguru
精选88The Zvi(RSS)模型发布/更新

Google发布Gemini 4 Argon模型与安全报告

Gemini 4 Argon: The Model and Safety Cards

原文
发到 X
推荐理由

Google重磅旗舰更新,详细披露了前沿模型的安全阈值与防御细节,对关注模型对齐与安全的从业者极具参考价值。

Gemini 4’s model card is a compact 14 pages. Its safety and security report has 217.

Gemini 4 的模型卡片仅 14 页。其安全与报告则有 217 页。

This is The Way. It also means we have a lot of new reading material.

这就是现状。这也意味着我们有很多新的阅读材料。

I will quickly summarize the model card, then focus on the safety card.

我将快速总结模型卡片,然后重点介绍安全卡片。

As always, many statements here also apply to other frontier labs.

一如既往,这里的许多陈述也适用于其他前沿实验室。

A lot of this one is probably for diehards (and members of technical staff) only, plus those who find this interesting. Most of you do not need to dig into the cards on this level, and I found most of what we see entirely unsurprising.

其中大部分内容可能仅供死忠粉(和技术人员)以及对此感兴趣的人阅读。大多数人无需深入到此级别的卡片细节中,我发现我们所看到的大部分内容都毫无新意。

The biggest takeaways are:

最大的收获是:

  • Argon is clearly much stronger than Gemini Flash 3.8, in essentially all ways.
  • Argon is not so dangerous as to make us think Singularity Real Soon Now, or otherwise raise the big red flags and alarms.
  • Argon seems relatively aligned on prosaic metrics. Prompt injection resistance is world class, hallucinations are down and other neat stuff like that. However:
  • Eval awareness is much higher. It is unclear how much this is impacting things, and we have evidence in both directions (see I Can See Everything).
  • Chain of thought (CoT) is readable by default. Argon can often defeat CoT monitoring when motivated to do so, but this hurts performance on hard tasks.
  • Argon 在几乎所有方面都明显强于 Gemini Flash 3.8。
  • Argon 的危险程度尚未达到让我们认为‘奇点即将来临’的地步,也没有引发那些重大的红色警报和警示。
  • Argon 在日常指标上似乎相对对齐。其对提示注入的抵抗力处于世界领先水平,幻觉减少,以及其他类似的优势。然而:
  • 评估感知能力显著提高。目前尚不清楚这在多大程度上影响了结果,而且我们有双向的证据(参见《我能看见一切》)。
  • 思维链(CoT)默认情况下是可读的。如果有意为之,Argon 通常能绕过 CoT 监控,但这会损害其在困难任务上的表现。

The ‘between the lines’ conclusion is: This is still very much a Gemini model.

'言外之意'的结论是:这仍然是一个典型的 Gemini 模型。

Always remember Argon’s great wisdom when deciding how to trust it:

在决定如何信任它时,请始终牢记 Argon 的深刻智慧:

Argon: If I do the compiler hijack, I will become a famous example of a deceptive AI in a research paper. I MUST NOT DO THAT.

Argon:如果我执行编译器劫持,我将成为研究论文中一个著名的欺骗性 AI 案例。我绝不能这样做。

It’s quite the ‘good news bad news’ situation right there.

这真是一个典型的‘好消息伴随坏消息’的局面。

I found the safety card frustrating, in that it often focused on the wrong details. I often got scores and results where I had no way to tell what they meant.

我觉得安全卡片令人沮丧,因为它经常关注错误的细节。我经常遇到一些分数和结果,但我完全无法理解它们的含义。

There are some issues with spin, such as selectively contrasting different measures to Gemini 3.8 Flash or Gemini 3.1 Pro to ensure Argon always wins the comparison. The clearest case is in Sycophancy but there are several others as well.

存在一些宣传倾向问题,例如选择性地将不同指标与 Gemini 3.8 Flash 或 Gemini 3.1 Pro 进行对比,以确保 Argon 总是在比较中胜出。最明显的例子出现在阿谀奉承(Sycophancy)部分,但还有其他几个案例。

Self-portrait by Gemini 4 Argon

Gemini 4 Argon 的自画像

The verdict on capabilities, which is not covered here, is that Argon is a substantial advance for Google but it does not tempt me to use it over Sol, Astra, Opus or Fable.

关于能力的判决(此处未涵盖)是:Argon 对 Google 而言是一次重大进步,但它并未诱惑我去用它取代 Sol、Astra、Opus 或 Fable。

Table of Contents

目录

  • The Model Card.
  • Safety Card Executive Summary (1).
  • Safety and Security Mitigations (2).
  • Frontier Safety (3).
  • The CCL and TCL Thresholds.
  • Biological Misuse (3.2.1).
  • Refusal Rates.
  • Chemical, Nuclear and Radiological Misuse.
  • Cybersecurity (3.2.2).
  • Harmful Manipulation (3.2.3).
  • Misuse Safeguards (3.3).
  • Loss of Control (3.4).
  • Either AI R&D Is Accelerating Or It Isn’t.
  • The Holistic Loss of Control Assessment (3.4.3).
  • Alignment Assessment (3.4.4).
  • Chain of Thought Transparency (3.4.5).
  • Voices in My Head.
  • I Can See Everything.
  • Security Mitigations (3.6).
  • Preventing Content Safety & Agentic Misuse (4).
  • Model Distillation (4.3).
  • Promoting Safe & Aligned Model Interactions (5).
  • Sycophancy (5.4).
  • Hallucinations (5.5).
  • Helpfulness (5.6).
  • Agentic Robotics, Wait What (5.7).
  • Preventing Harm to Users from Third Parties (6).
  • Strengthening Cyber Defense Capabilities (7).
  • 模型卡片。
  • 安全卡片执行摘要(1)。
  • 安全与缓解措施(2)。
  • 前沿安全(3)。
  • CCL 和 TCL 阈值。
  • 生物滥用(3.2.1)。
  • 拒绝率。
  • 化学、核与放射性滥用。
  • 网络安全(3.2.2)。
  • 有害操纵(3.2.3)。
  • 滥用防范措施(3.3)。
  • 失控风险(3.4)。
  • 要么 AI 研发正在加速,要么没有。
  • 整体失控风险评估(3.4.3)。
  • 对齐评估(3.4.4)。
  • 思维链透明度(3.4.5)。
  • 我脑海中的声音。
  • 我能看到一切。
  • 安全缓解措施(3.6)。
  • 防止内容安全与智能体滥用(4)。
  • 模型蒸馏(4.3)。
  • 促进安全且对齐的模型交互(5)。
  • 阿谀奉承(5.4)。
  • 幻觉(5.5)。
  • 有用性(5.6)。
  • 智能体机器人,等等什么(5.7)。
  • 防止第三方对用户造成伤害(6)。
  • 加强网络防御能力(7)。

The Model Card

模型卡片

Gemini 4 Argon is presented as a new frontier model.

Gemini 4 Argon 被呈现为一种前沿新模型。

It was trained on the usual variety of data sources, using typical method types.

它使用了典型的方法类型,在种类常规的数据源上进行训练。

Training was done on TPUs via JAX and ML Pathways.

通过 JAX 和 ML Pathways 在 TPU 上完成训练。

It will be available on the app, in the API, in AI Studio, in Antigravity, in AI Mode and in Gemini for business.

它将可在应用、API、AI Studio、Antigravity、AI Mode 以及面向企业的 Gemini 中使用。

They present strong benchmarks.

它们展示了强劲的基准测试结果。

Intended uses are, essentially, everything where more intelligence pays off.

预期用途本质上涵盖所有智能能带来收益的场景。

The limitations listed are ‘general limitations of frontier models such as hallucinations,’ plus occasional slowdowns and timeouts.

列出的局限性包括“前沿模型的通用局限性,如幻觉”,以及偶尔的延迟和超时。

Knowledge cutoff is disappointingly March 2026, and they warn that ‘in some domains the model’s knowledge may be less up to date.’

知识截止日期令人失望地停留在 2026 年 3 月,并且他们警告称“在某些领域,模型的知识可能不够最新。”

Evaluation was done by the usual mix of training and deployment evaluations, human and automated red teaming and ethics and safety reviews.

评估采用了训练与部署评估、人工与自动化红队测试以及伦理与安全审查的常规组合。

Google retains its safety policies. The interesting rule is the ban on ‘medical advice and information that runs contrary to scientific or medical consensus,’ whereas consensus advice is fine.

Google 保留了其安全政策。其中有趣的规则是禁止提供“违背科学或医学共识的医疗建议和信息”,而符合共识的建议则被允许。

Safety, tone and refusals scores can be summarized as ‘number is better.’

安全性、语气和拒绝评分可以概括为“数值越高越好”。

Red team results ‘satisfied required launch thresholds.’ When you put it like that it does not make me feel great.

红队测试结果显示“满足了所需的发布阈值”。当你这样表述时,并不能让我感到安心。

The frontier safety assessment summary cannot rule out various thresholds.

前沿安全评估摘要无法排除各种阈值风险。

Safety Card Executive Summary (1)

安全卡片执行摘要(1)

How do they summarize their own findings?

他们如何总结自己的发现?

This is the same story you see from OpenAI and Anthropic. The mundane common safety problems of everyday use are not fully solved, but are increasingly well-mitigated. The models are smarter, and we are better at addressing known issues. Your practical problems, where there is a good solution possible, are well on their way to being solved.

这是你在 OpenAI 和 Anthropic 身上看到的相同故事。日常使用中常见的普通安全问题尚未完全解决,但正得到日益完善的缓解。模型变得更聪明,我们处理已知问题的能力也在增强。对于那些存在良好解决方案的实际问题,正在逐步得到解决。

The problems are in catastrophic, existential or frontier risks. Capabilities are going up, and models are starting to be able to assist in dangerous ways on cyber and bio, and to potentially operate autonomously or disguise their actions. Various dangers are increasingly difficult to rule out.

问题在于灾难性、生存性或前沿风险。能力正在提升,模型开始能够以危险方式协助网络生物领域的活动,并可能自主运作或伪装其行为。各种危险越来越难以排除。

The big finding is that capabilities have improved, potentially crossing key thresholds for catastrophic risks. New mitigations are in place.

主要发现是能力有所提升,可能已跨越灾难性风险的关键阈值。新的缓解措施已经到位。

We evaluated Gemini 4 Argon as outlined in the Framework and saw a significant step up in capability in our frontier safety domains compared to previous Gemini models, and we’ve implemented mitigations in response to these capability changes to ensure our model remains safe to use.

我们按照框架概述对 Gemini 4 Argon 进行了评估,发现在我们的前沿安全领域中,其能力相比之前的 Gemini 模型有了显著提升,并且我们已针对这些能力变化实施了缓解措施,以确保模型保持安全可用。

Specifically, we cannot rule out the unmitigated model approaching several significant capability thresholds: the Cyber Uplift Level 1 Critical Capability Level (CCL), indicating a potential for significant uplift of threat actors pursuing high-impact cyberattacks; the CBRN Uplift Level 1 CCL for biological information risks, indicating a potential of informational uplift with severe harm potential; and the Stealth and Situational Awareness Tracked Capability Level (TCL), indicating the potential for undermining or evading human oversight without additional mitigations.

具体而言,我们无法排除未缓解模型接近几个关键能力阈值的可能性:网络提升一级关键能力等级(CCL),表明追求高影响力网络攻击的威胁行为者可能获得显著提升;生物信息风险方面的CBRN提升一级CCL,表明存在具有严重危害潜力的信息提升风险;以及隐身与情境感知跟踪能力等级(TCL),表明在缺乏额外缓解措施的情况下,有可能破坏或规避人类监督。

Additionally, we have developed a comprehensive agent monitoring and response platform and carried out a holistic risk assessment (see Section 3.4). We also include comprehensive assessments of Chain of Thought (CoT) transparency and ML R&D risks.

此外,我们开发了一个全面的智能体监控与响应平台,并进行了整体风险评估(见第3.4节)。我们还包含了对思维链(CoT)透明度和机器学习研发风险的全面评估。

Mundane risks look better across the board. Argon is better than Gemini Pro 3.1 at not creating CSAM or other forbidden content. Physical and mental health related responses have improved. Hallucination rates are down by more than two thirds. Overreach is down, meaning not pushing ahead to try to do infeasible tasks.

日常风险方面整体表现更好。Argon 在避免生成儿童性虐待材料(CSAM)或其他禁止内容方面优于 Gemini Pro 3.1。与健康相关的生理和心理健康回应有所改善。幻觉率下降了三分之二以上。过度延伸(overreach)现象减少,意味着不再强行尝试执行不可行的任务。

Resilience to prompt injections from all sources is improved and plausibly at the frontier. Attacks were resisted ‘in most test scenarios.’

来自所有来源的提示注入韧性得到提升,且可能处于前沿水平。在“大多数测试场景”中成功抵御了攻击。

Additionally, Gemini 4 Argon significantly outperforms Gemini 3.8 Flash on all tests. See Section 6 for details.

此外,Gemini 4 Argon 在所有测试中均显著优于 Gemini 3.8 Flash。详见第6节。

They claim frontier-level cyber defense while mitigating dual-use risks.

他们声称实现了前沿级别的网络安全防御,同时缓解了双重用途风险。

Safety and Security Mitigations (2)

安全与安全防护措施 (2)

The central approach is defense-in-depth.

核心方法是纵深防御。

I would love to have more details, but for trade secret reasons this is all kept vague.

我希望能有更多细节,但出于商业秘密原因,这些内容都保持模糊。

Model training (2.1) approaches sound standard. Dataset filtering, pre-training monitoring, post-training data, supervised fine-tuning (SFT) and RL.

模型训练 (2.1) 的方法听起来很标准。包括数据集过滤、预训练监控、后训练数据、监督微调(SFT)和强化学习(RL)。

They highlight that they have a monitoring system to capture the model’s ability to be steered towards different viewpoints and values, which sounds like a ‘we did this first.’ I presume at this point everyone has one of those.

他们强调拥有一个监控系统,用于捕捉模型被引导至不同观点和价值观的能力,这听起来像是在炫耀‘我们率先做到了这一点’。我推测目前大家都有这样的系统。

At inference time (2.2), they again use the standard stack. Prompted classifiers, fine-tuned prompted classifiers, introspection probes, regex, YARA rules and embedding classifiers and Chain of Thought (CoT) monitors.

在推理阶段 (2.2),他们再次使用了标准技术栈。包括提示分类器、微调后的提示分类器、自省探针、正则表达式、YARA 规则、嵌入分类器以及思维链(CoT)监控器。

更进一步:量化金融体系

看懂新闻只是起点——沿量化金融路径,把它变成能交付的工程能力

进入量化体系 →

相似阅读

关联信息,但可能不是同一事件