Google于当地时间10月28日正式发布新一代前沿人工智能模型Gemini 4 Carbon，重点提升软件编程、智能体执行复杂任务以及网络安全防御能力。Google称，新模型在发现软件漏洞、验证漏洞可利用性以及修复安全缺陷等多个环节取得显著进步，并在多项基准测试中领先于OpenAI和Anthropic的旗舰模型。不过，考虑到其强大的网络安全能力可能带来的潜在风险，Google目前仅向部分经过筛选的网络安全防御机构开放使用，暂时不会立即向所有开发者和普通用户提供。
Gemini 4 Carbon是Google本月推出Gemini 4 Argon之后的又一款前沿模型。Google表示，新模型在Argon的基础上进一步强化了代码生成、智能体工作流和长时间复杂任务处理能力，尤其针对网络安全领域进行了重点优化。公司希望通过提升人工智能发现和修复软件漏洞的能力，帮助企业更有效地保护系统，同时降低新代码引入安全缺陷的风险。
在综合性能方面，Google援引Artificial Analysis发布的Artificial Analysis Intelligence Index 5.0评测结果称，Gemini 4 Carbon以59分排名第一，领先于Anthropic的Claude Opus 5.5和OpenAI的GPT-6 Astra。该榜单综合考察多个维度的模型能力，但不同基准测试的任务类型和评估方法存在差异，因此综合排名并不意味着模型在所有实际应用中都具有优势。
网络安全是此次发布的核心。Google公布的数据显示，Gemini 4 Carbon在SEC-bench Pro基准测试中取得88.2%的成绩，该测试主要评估模型发现真实软件漏洞的能力；在ExploitGym测试中取得54.8%的成绩，用于衡量模型验证漏洞可利用性、构建相应利用方案的能力；在CWE-bench测试中则取得71%的成绩，重点评估模型能否针对已识别的安全弱点提出有效修复方案。
这些能力覆盖了软件安全工作中的多个关键环节。传统安全工具通常依赖预先定义的规则、已知漏洞特征和特定测试流程，而先进人工智能模型有望进一步分析复杂代码，发现隐藏在大型软件系统中的缺陷，并协助开发者理解漏洞形成的原因，最终生成针对性补丁。
不过，漏洞发现与漏洞利用之间存在重要区别。能够识别程序中的缺陷，并不意味着模型一定能够在真实环境中成功利用这些缺陷；同样，生成一段看似合理的修复代码，也不代表它已经彻底消除漏洞。Google公布的测试成绩反映的是特定评测条件下的表现，实际安全效果仍取决于代码环境、测试方法以及后续人工验证。
Google还表示，Gemini 4 Carbon在内部测试中生成安全代码的成功率超过Gemini 4 Argon的两倍。Chrome团队也报告称，在其内部评估中，Carbon生成高质量安全补丁的表现明显优于某些其他实验室的前沿模型。不过，这些结果主要来自Google自己的测试环境，具体评估条件和测试任务可能影响最终表现，仍需要更多独立测试加以验证。
除网络安全之外，Gemini 4 Carbon还针对需要长时间连续执行的工程任务进行了优化。Google表示，内部团队已经观察到模型能够连续运行数天，在经历多次失败和尝试之后，仍继续处理复杂的多步骤工程问题。
这类任务可能涉及大型代码库中的疑难错误定位、低层系统问题排查，以及跨多个模块实施代码修改。与只需回答单个问题的聊天机器人不同，智能体需要制定计划、调用工具、检查结果，并根据执行过程中的反馈调整方案。任务越复杂，模型就越需要保持上下文连贯性，并避免在遇到困难时过早放弃。
Google称，Carbon在处理此类长期任务时，更善于识别失败路径，避免为了完成任务而采取不可靠的捷径。在实际开发过程中，这有望帮助工程师减少重复排查的时间，让人工智能承担更多持续性的软件维护和调试工作。
与此同时，Google强调，新模型在可靠性方面也有所改善。公司表示，Carbon相较Argon将未经证实的断言减少了约一半。面对证据不足或存在歧义的问题，模型更倾向于进一步检索资料、指出不同信息之间的矛盾，或者明确说明哪些步骤仍然无法完成，而不是直接给出缺乏依据的答案。
在提示注入攻击防御方面，Google公布的内部评估结果同样值得关注。提示注入攻击通常通过将恶意指令隐藏在网页、文件或其他外部内容中，诱导人工智能系统偏离原本的任务目标。Google表示，Carbon在Gray Swan的间接提示注入测试中，攻击成功率从Argon的0.7%进一步降至0%，即在该测试覆盖的攻击尝试中，没有出现成功案例。
不过，单一基准测试中未观察到攻击成功，并不能证明模型能够抵御所有提示注入攻击。真实环境中的攻击方式可能不断变化，模型还需要结合权限管理、工具调用限制、输入内容审查以及人工监督等措施，才能构建更可靠的安全体系。
正因为Carbon在网络安全领域的能力进一步增强，Google决定采取分阶段开放策略。新模型首先通过公司的Fairwind计划提供给部分经过筛选的网络安全防御机构。Google表示，公司正在参与美国政府推动的前沿模型发布前自愿测试流程，并计划根据首批用户反馈持续调整安全防护机制，再逐步向更多可信合作伙伴开放。
Google尚未公布Carbon面向所有开发者、企业和普通用户开放的确切日期。公司表示，希望在完成必要的测试和防护措施调整后，尽快将模型引入更多产品和服务。
这种有限开放策略反映出人工智能行业面临的一项现实挑战：能够帮助安全团队发现漏洞的模型，同样可能被恶意人员用于寻找攻击入口或开发利用工具。模型能力越强，企业越需要评估其在不同使用场景下可能带来的双重用途风险。
在价格方面，Gemini 4 Carbon的初始定价与Gemini 4 Argon相同，每百万输入Token收费2美元，每百万输出Token收费10美元。Google还为缓存输入Token提供大幅折扣，折扣幅度为输入Token价格的95%。不过，这一价格属于推广期定价，Google公布的后续价格为每百万输入Token 4美元、每百万输出Token 20美元，正式调整时间尚未明确。
从产品定位来看，Carbon并不是一款单纯追求聊天体验的新模型，而是更侧重于复杂软件工程、持续执行任务以及网络安全防御的专业工具。它的价值不仅在于能否给出正确答案，还在于能否在较长时间内持续完成工作，并在遇到错误、模糊要求或潜在安全风险时作出恰当处理。
此次发布也显示，Google正在加快前沿模型的迭代节奏。Gemini 4 Argon推出后不久，Carbon便进一步提升了多个关键能力指标。对于企业用户而言，这种快速迭代有望缩短人工智能技术进入实际软件开发和安全运营流程的时间，但也增加了独立评测、持续监控和安全验证的重要性。
目前，Gemini 4 Carbon在多项由Google公布的基准测试中取得了领先成绩，但其真实世界表现仍有待更多独立机构和实际用户检验。随着网络攻击手段日益复杂，人工智能既可能成为安全团队的重要防御工具，也可能增加漏洞发现和利用的自动化程度。Google选择先向经过筛选的防御机构开放模型，正是希望在释放技术能力的同时，进一步评估其风险，并逐步建立相应的使用边界。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Google officially launched the new generation of publish front-line artificial intelligence model Gemini 4 Carbon on October 28, local time, focusing on upgrading software programming, AI agent on complex missions and cyber-security defence capabilities. According to Google, the new model has made significant progress in the identification of gaps in software, the availability of validation gaps and the repair of security deficiencies, and has led to flagship models of Open AI and Anthropic in various benchmark tests. However, in view of the potential risks that its strong network security capability may pose, Google is currently open to use only by partially screened cybersecurity defences and will not be available immediately to all developers and ordinary users.
Gemini 4 Carbon is another cutting-edge model that Google launched this month after Gemini 4 Argon. According to Google, the new model further enhanced code generation, AI agentworkflow and long-term complex mission processing capabilities based on Argon, with a particular focus on the area of cybersecurity. By improving the ability of artificial intelligence to detect and repair software gaps, companies want to help enterprises protect systems more effectively while reducing the risk of new codes introducing security deficiencies.
In terms of integrated performance, Google cites Artificial Analysis Inteligence Index 5.0 of Artificial Analysis publish, according to which Gemini 4 Carbon ranks 59th, ahead of Claude Opus 5.5 of Anthropic and GPT-6 Astra of Open AI. The list takes a comprehensive look at model capabilities at multiple dimensions, but there are differences in task types and assessment methods for different baseline tests, so the composite ranking does not mean that the model has advantages in all practical applications.
Cybersecurity is at the heart of publish this time. The data published by Google show that Gemini 4 Carbon achieved 88.2 per cent in the SEC-bench Pro benchmark test, which primarily assessed the ability of the model to detect real software gaps; 54.8 per cent in the Expluit Gym test, which measured the availability of the model to validate gaps and build capacity to use the programme accordingly; and 71 per cent in the CWE-bench test, which focused on the ability of the model to propose effective rehabilitation programmes for identified security vulnerabilities.
These capabilities cover several key aspects of software security. While traditional security tools often rely on predefined rules, known loophole characteristics and specific testing processes, advanced artificial intelligence models are expected to further analyse complex codes, detect deficiencies hidden in large software systems and assist developers in understanding the causes of loopholes and eventually generate targeted patches.
However, there are important differences between the discovery of loopholes and their utilization. The ability to identify defects in the process does not mean that the model must be able to use them successfully in the real environment; similarly, the generation of a seemingly reasonable repair code does not mean that it has completely eliminated the loophole. The results of the tests published by Google reflect the performance under certain assessment conditions, and the actual safety effects continue to depend on the code environment, testing methods and subsequent manual validation.
Google also states that Gemini 4 Carbon generated security codes twice as many Z success rates as Gemini 4 Argon in internal tests. The Chrome team also reported that, in its internal assessment, Carbon produced high-quality safety patches that performed significantly better than front-line models in some other laboratories. However, these results are mainly from Google ' s own testing environment, and specific assessment conditions and testing missions may affect final performance, which still require additional independent testing to validate.
In addition to network security, Gemini 4 Carbon has been optimized for engineering tasks that require continuous work over a long period of time. Google states that the internal team has observed that the model can run for several days and continues to address complex multi-step engineering issues after repeated failures and attempts.
Such tasks may involve problematic error positioning in large code libraries, the sorting of low-level system problems, and code changes across multiple modules. Unlike chat robots who need only answer a single question, AI agent needs to develop plans, call tools, check results and adjust programmes based on feedback during implementation. The more complex the task, the more necessary it is to maintain context continuity and avoid premature abandonment in the face of difficulties.
According to Google, in dealing with such long-term tasks, Carbon is better at identifying failure paths and avoiding unreliable shortcuts in order to complete the task. This is expected to help engineers reduce the time for repeat searches and allow artificial intelligence to undertake more ongoing software maintenance and commissioning work during the actual development process.
At the same time, Google stressed that the new model had also improved its reliability. According to the company, carbon reduced unsubstantiated assertions by about half compared to Argon. Faced with problems of insufficient evidence or ambiguity, the model prefers to search further for information, to point out contradictions between different information, or to specify which steps are still not complete, rather than directly providing unfounded answers.
The results of the internal assessment published by Google are also of concern in terms of alerting to attack defences. Injecting an attack usually induces artificial intelligence systems to deviate from their original mission targets by hiding malicious instructions in web pages, files or other external content. According to Google, no success was reported in the attack on success rate, which was further reduced from 0.7 per cent to 0 per cent in Argon, during the indirect infusion test of Gray Swan.
However, the failure to observe the success of the attack in the single benchmark test does not prove that the model is able to withstand all tips of injection. The pattern of attack in the real environment may change and the model will need to be combined with measures such as permission management, tool call restrictions, input content review and manual monitoring to build a more reliable security system.
As Carbon ' s capacity in the area of cybersecurity was further enhanced, Google decided to adopt a phased open approach. The new model was first made available to selected cybersecurity defence agencies through the company ' s Fairwind plan. Google indicated that it was participating in the United States Government-driven process of voluntary pre-testing of the front-line model publish and planned to adapt the security protection mechanism on a continuous basis, based on feedback from the first users, to be gradually opened to more credible partners.
Google has not yet published the exact date of Carbon's opening to all developers, businesses and ordinary users. The company expressed the hope that the model would be introduced into more products and services as soon as the necessary tests and protective measures were completed.
This limited open-door strategy reflects a real challenge for the artificial intelligence industry: models that can help security teams to detect loopholes may also be used by malicious individuals to find entry points or develop tools to use them. The greater the modelling capacity, the greater the need for enterprises to assess the dual-use risks that they may pose in different contexts of use.
In terms of prices, Gemini 4 Carbon's initial pricing was the same as Gemini 4 Argon, with Token charged $2 per million and Token $10 per million. Google also provided a substantial discount for cache to enter Token, which was 95 per cent of Token ' s price. However, this price is the extension period price, with Google publishing a subsequent price of $4 per million for Token and $20 per million for Token, the official timing of the adjustment is not yet clear.
From the point of view of product positioning, Carbon is not a new model of purely chat-seeking experiences, but rather a professional tool that focuses more on complex software engineering, ongoing missions and cyber-security defences. Its value lies not only in the ability to give the right answer, but also in the ability to carry out the work over a longer period of time and to properly address mistakes, vague requirements or potential security risks.
This time publish also shows that Google is accelerating the iterative pace of the frontier model. Shortly after Gemini 4 Argon was launched, Carbon further enhanced several key capacity indicators. For enterprise users, this rapid iterativeity is expected to reduce the time that artificial intelligence technology takes into actual software development and secure business processes, but also increase the importance of independent evaluation, continuous monitoring and security validation.
Gemini 4 Carbon is currently leading in many of the benchmarking tests published by Google, but its real world performance still needs to be tested by more independent institutions and actual users. As cyberattack methods become more sophisticated, artificial intelligence can become both an important defence tool for security teams and an increased degree of automation in the detection and use of loopholes. Google chose to open itself first to a selected defence structure, hoping to release its technical capabilities while further assessing its risks and gradually establishing the corresponding use boundaries.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.