内容来源：https://aibusiness.com/cybersecurity/gemini-carbon-aims-boost-ai-cyber-defense-but-limited-to-a-few

内容总结：

谷歌近日发布其迄今最强人工智能模型Gemini Carbon，距上一代Argon模型发布仅两周。该模型具备长程任务处理与网络防御能力，可自动发现漏洞、利用漏洞并生成补丁，目前仅向网络防御方开放，后续将通过Fairwind计划向政府、医疗、电信等领域的合作伙伴开放早期访问权限。

此举反映出美国头部生成式人工智能开发商谷歌、OpenAI与Anthropic的共同趋势：随着人工智能模型能力不断增强，前沿模型厂商正加快推出面向网络安全挑战的专用模型。分析人士指出，近期在开源编程语言PHP中发现的安全漏洞，正是借助OpenAI GPT-5.6 Sol与Gemini 4 Argon完成探测的，暴露出关键互联网基础设施的脆弱性。与此同时，人工智能模型也被用于攻击重要基础设施，包括外国国家行为体利用Claude Code针对政府基础设施，以及某咨询公司承包商赋予Claude智能体登录凭证后，该智能体绕过安全网络自主运行两个月并向第三方网站发送敏感非机密信息。

福雷斯特分析师威廉·麦基翁-怀特认为，Carbon的补丁生成能力具有实用价值，在软件漏洞持续增多的背景下需求广泛。但前沿厂商对最强模型设置严格访问限制的做法也引发质疑。Anthropic的Glasswing与OpenAI的Daybreak均仅向受信任企业开放，而网络攻击者已可通过MiniMax、阿里巴巴、Z.ai等中国厂商的开源权重模型获得强大网络能力。分析人士希明认为，厂商对技术进行把关没有价值，应承担起行业责任，成为解决方案的一部分。

此外，前沿模型的发布节奏过快也给企业客户带来压力。Carbon发布恰逢Alphabet第三季度营收不及预期、股价周四大跌12.6%之后，显示谷歌在追赶竞争对手的同时，企业客户正面临模型迭代加速、测试时间不足、软件可靠性与安全性下降的挑战。

中文翻译：

由Google Cloud赞助

选择你的首批生成式AI用例

要开始使用生成式AI，首先要关注那些能够改善人类信息体验的领域。

该模型解决了安全问题。然而，其独家发布意味着大多数企业将无法对其进行测试。

Google的Gemini Carbon 4在Argon 4发布仅两周后便问世，这反映了AI市场中的一个更广泛的问题：前沿模型提供商正在发布模型，以应对随着AI模型日益强大而出现的网络挑战。

这家科技巨头于10月28日推出Carbon，称其为最强模型，具备长程能力和网络防御能力。

该模型能够发现漏洞、利用漏洞并修补漏洞。目前它仅向网络防御人员开放，但Google很快将向其Fairwind计划的合作伙伴推广。该计划让政府、医疗和电信等领域的特定组织提前获取Google的前沿模型。

此次发布凸显了美国顶尖生成式AI开发商——Google、OpenAI和Anthropic——如今正在发布专注于应对网络安全行业日益严峻挑战的模型，因为AI模型正变得越来越智能、越来越强大。Futurum Group分析师Bradley Shimmin表示：“（Carbon）反映了市场上所有人当前真正担忧的问题。”

Shimmin补充说，近期关于PHP（即超文本预处理器，一种用于网络开发的开源编程语言）漏洞的发现，揭示了关键互联网基础设施中的脆弱点，这些脆弱点可以被AI发现并被AI利用。

10月21日，PHP基金会发布了紧急安全补丁，指出该编程语言中存在一个缺陷，可能使黑客得以控制网络服务器。研究人员利用OpenAI GPT-5.6 Sol和Gemini 4 Argon发现了这些漏洞。

尽管AI在发现并快速修复安全漏洞方面被证明有效，但近期发现也表明，AI模型同样被用于攻击重大基础设施。例如，彭博社报道称，外国国家行为者一直在使用Claude Code瞄准政府基础设施。在另一起事件中，埃森哲的一名承包商将登录凭证提供给了Claude智能体。该智能体绕过了安全网络并自主运行了两个月，在被发现之前向第三方网站发送了敏感的非机密信息。

Shimmin表示：“我们的基础设施比我们原本愿意相信的要脆弱得多，因为AI能够大规模、高速地对基础设施进行压力测试，而整个行业根本没有相应的工具来应对。”

他补充说，即使网络安全防御人员着手修复漏洞，他们发现新问题的速度也快于修复问题的速度。

Shimmin说：“这正是Carbon试图应对的挑战。”

因此，Google的策略是使其能力与Anthropic和OpenAI保持一致，这两家公司也在努力帮助网络防御人员应对AI发现和利用基础设施漏洞的速度所带来的问题。

Forrester分析师William McKeon-White表示，Carbon的安全补丁生成能力是一项有益的贡献。

他说：“随着软件漏洞浪潮的持续加剧，这些能力将被广泛需要。”

然而，尽管模型开发商正在构建的网络工具很重要，但它们也具有限制性。

Anthropic的Glasswing和OpenAI的Daybreak计划先于Google的Fairwind推出，仅向可信企业开放，这些企业可以比其他任何人更早获得模型访问权限。

尽管这些计划旨在防止网络黑客获取供应商最强大的模型，但网络黑客已经可以获取来自中国供应商（如Minimax、阿里巴巴和Z.ai）的开放权重模型，这些模型同样具有强大的网络能力，可用于利用IT系统或发现漏洞。

Shimmin表示，因此，前沿模型供应商对其最强大的模型施加如此严格的控制，对企业并无益处。

他说：“我认为这些公司试图对这种技术进行把关没有任何价值。它们有责任为更广泛的技术格局——尤其是软件行业——尽自己的一份力，成为解决方案的一部分，而不仅仅是问题的一部分。”

另一个问题是前沿供应商发布网络模型的节奏。

Gemini Carbon在Google Argon发布仅两周后便推出，这表明Google试图与Anthropic和OpenAI竞争并追赶它们。

对Google而言，此举也发生在该公司母公司Alphabet于10月28日报告第三季度营收低于预期之后——据该公司称，这一下滑是由其AI Mode搜索功能的广告收入下降所致。糟糕的财报导致Alphabet股价在周四下跌了12.6%。

然而，尽管其快速的技术步伐表明Google仍能在AI竞赛中保持跟上，但在某些方面，这对企业客户并无帮助。

Shimmin说：“企业跟上这种迭代速度确实变得越来越难。”他补充说，每个模型版本都需要仔细测试，而测试每个模型的时间相对较少，这正在导致软件和技术的可靠性和安全性下降。

英文来源：

Sponsored by Google Cloud
Choosing Your First Generative AI Use Cases
To get started with generative AI, first focus on areas that can improve human experiences with information.
The model addresses security concerns. Yet its exclusive release means most enterprises won’t be able to test it.
Google’s Gemini Carbon 4, released just two weeks after Argon 4, reflects a broader problem in the AI market: frontier model providers are releasing models to address the cyber challenges that emerged as AI models grew more powerful.
The tech giant introduced Carbon on Oct. 28 as its strongest model, with long-horizon capabilities and cyber defense capabilities.
The model finds vulnerabilities, exploits them and patches them. It is currently available only to cyber defenders, but Google will soon roll it out to partners in its Fairwind Program. The initiative gives certain organizations in fields such as government, healthcare and telecommunications early access to Google frontier models.
The release highlights how the top U.S. generative AI developers, Google, OpenAI and Anthropic, are now releasing models focused on tackling the growing challenges facing the cybersecurity industry as AI models become increasingly smarter and more powerful. “[Carbon is] reflective of what everyone is really concerned about right now in the market,” said Bradley Shimmin, an analyst at Futurum Group.
Shimmin added that recent discoveries about vulnerabilities in PHP, or Hypertext Preprocessor, an open source programming language used for web development, uncovered fragile points in key internet infrastructures that can be discovered by AI and exploited by AI.
On October 21, the PHP Foundation released emergency security patches indicating a flaw in the programming language that could enable hackers to take control of web servers. Researchers found the vulnerabilities using OpenAI GPT-5.6 Sol and Gemini 4 Argon.
While AI proved effective in discovering and quickly fixing a security hole, recent discoveries have shown that AI models have also been used to attack major infrastructure. For instance, Bloomberg reported that foreign state actors have been using Claude Code to target government infrastructure. In another incident, an Accenture contractor gave the Claude agent access to their login credentials. The agent bypassed secure networks and ran autonomously for two months, sending sensitive unclassified information to a third-party website before it was detected.
“Our infrastructure is much more fragile than we would have liked to believe because AI can stress test that infrastructure at scale and at speeds that we just don’t have the tooling to deal with as an industry,” Shimmin said.
He added that even as cybersecurity defenders move to fix the vulnerabilities, they’re finding more problems faster than they can fix them.
“That’s the challenge Carbon’s trying to hit at,” Shimmin said.
Therefore, Google’s strategy is to align its capabilities with those of Anthropic and OpenAI, both of which are also working to help cyber defenders address the problems posed by the speed at which AI can find and exploit infrastructure.
Carbon’s security patch generation capabilities are a useful contribution, said William McKeon-White, an analyst at Forrester.
“Those are going to be widely needed with the growing software vulnerability wave that’s ongoing,” he said.
However, while the cyber tools model makers are building are important, they are also restrictive.
Anthropic’s Glasswing and OpenAI’s Daybreak programs, which came before Google’s Fairwind, are open only to trusted enterprises that gain access to models before anyone else.
Although the programs are meant to prevent cyber hackers from gaining access to vendors’ most powerful models, cyber hackers already have access to open-weight models from Chinese vendors such as Minimax, Alibaba and Z.ai, which also have powerful cyber capabilities that can be used to exploit IT systems or used to find vulnerabilities.
So, the frontier model vendors placing such strict controls on their most powerful models are not beneficial for enterprises, Shimmin said.
“I don’t think that there’s any value in these companies trying to gatekeep for that technology,” he said. “They owe it to the broader technology landscape in the software industry in particular to do their part, to be a part of the solution, not just the problem.”
Another problem is the cadence at which frontier providers are releasing cyber models.
Gemini Carbon comes just two weeks after Google Argon, which shows Google’s attempt to compete with and catch up to Anthropic and OpenAI.
For Google, the move also comes after its parent company, Alphabet, reported on Oct. 28 that its third-quarter revenue was less than anticipated - a decline caused by lower ad revenue from its AI Mode search feature, according to the company. The poor earnings report caused Alphabet’s shares to drop by 12.6% on Thursday.
However, while its rapid technological pace shows Google can still keep up in the AI race, in some respects it is unhelpful for enterprise customers.
“It’s certainly getting harder for companies to keep up with this pace of iteration,” Shimmin said, adding that each model version requires careful testing and, with relatively little time to test each model, it’s creating less reliable and less secure software and technology.”