Anthropic 推出开源漏洞扫描服务 OSS Scanner,免费面向开源项目

来源:Anthropic:Research(发表成果 · 网页) 2026年10月7日 00:00 AIHOT 评分:69 精选
摘要:Anthropic 发布 OSS Scanner,一个可选加入的开源漏洞扫描服务,用其最强模型(包括 Claude Mythos)定期免费扫描开源项目,输出为全模型生成、无人工复核的报告。
推荐理由:原文给出扫描准确率、人工验证数据和维护者反馈,读者可评估这个免费开源安全扫描服务的实际信号质量与参与方式。
# Anthropic 推出开源漏洞扫描服务 OSS Scanner,免费面向开源项目 > 原标题:Launching an opt-in vulnerability-finding service for open-source software - 来源:Anthropic:Research - 发布时间:2026-10-07T16:00:00.000Z - AIHOT:https://aihot.news/items/xfch7s2npl8a4651n2ymips4k - 原文:https://www.anthropic.com/research/launching-opt-in-vuln-finding-service-for-open-source ## 摘要 Anthropic 发布 OSS Scanner,一个可选加入的开源漏洞扫描服务,用其最强模型(包括 Claude Mythos)定期免费扫描开源项目,输出为全模型生成、无人工复核的报告。 ## 推荐理由 原文给出扫描准确率、人工验证数据和维护者反馈,读者可评估这个免费开源安全扫描服务的实际信号质量与参与方式。 ## 正文 · 中文译文 _我们推出 [OSS Scanner](https://red.anthropic.com/oss-scanner),这是一款面向开源生态系统的可选加入漏洞扫描器,其设计源于我们在 [Project Glasswing](https://www.anthropic.com/glasswing) 期间使用 Claude 查找漏洞的经验。加入该项目的项目将免费获得由我们最强模型进行的全面、定期安全扫描。_ 正如我们在本博客上[大量报道](https://www.anthropic.com/research/zero-days)的那样,语言模型在[发现漏洞](https://www.anthropic.com/research/mythos-preview)方面的能力已迅速提升。在学术性漏洞发现基准 [CyberGym](https://arxiv.org/abs/2506.02548) 上,LLM 从去年年初只能发现不到 20% 的漏洞,提升到今年能发现超过 85% 的漏洞。因此,开源维护者从收到 LLM 发来的[大多是垃圾内容](https://daniel.haxx.se/blog/2025/07/14/death-by-a-thousand-slops/),转变为收到[高质量的漏洞报告](https://lore.kernel.org/linux-media/CAHk-=wi4zC+Ze8e+p3tMv8TtG_80KzsZ1syL9anBtmEh5Z40vg@mail.gmail.com/)。 在过去六个月里,我们使用最新模型扫描了一些世界上最重要的软件项目中的漏洞。我们发现了[超过 29,000 个候选漏洞](https://red.anthropic.com/2026/cvd/),但只能手动审查和分类其中约 6,000 个。虽然 6,000 个漏洞已经很多,但我们仍然受限于人工验证这些发现的能力。 我们正在努力扩大漏洞披露的规模。与此同时,收到我们首批报告的维护者越来越频繁地直接要求我们批量提交所有未经核实的报告及建议补丁:迄今为止,在维护者要求接收我们掌握的全部内容后,我们已直接向他们发送了近 5,000 份报告——即使这些报告尚未经过验证。由于漏洞利用现在可以在[几分钟内](https://www.anthropic.com/research/n-days)开发出来,能够更快发现并修复漏洞的项目,可以更好地保护其软件,抵御那些争相寻找并利用这些相同弱点的攻击者。 我们将继续通过现有的[协调漏洞披露(CVD)流程](https://www.anthropic.com/coordinated-vulnerability-disclosure)手动披露经人工验证的漏洞报告,尤其是针对那些没有资源自行分类报告的项目。但我们现在为那些希望一有漏洞报告就立即收到的项目提供可选的快速通道。 ## 我们的漏洞扫描器 受[Google 的 OSS-Fuzz](https://google.github.io/oss-fuzz/) 对开源生态系统积极影响的启发——该项目使用模糊测试器扫描开源软件中的漏洞——我们推出 [OSS Scanner](https://red.anthropic.com/oss-scanner),用我们最强的语言模型查找开源代码中的漏洞。我们的通用代码扫描与修补产品 [Claude Security](https://claude.com/product/claude-security) 侧重于帮助企业防御其系统,而 OSS Scanner 则为开源项目免费提供安全审计。 这款可选加入的漏洞扫描器的输出将完全由模型生成,不经过人工审查或分类。这将实现更快、更频繁的扫描,但也意味着报告有可能不正确或无效。这些报告将由我们最强的模型(包括 Claude Mythos)生成,以便为开源项目提供最大的防御优势。 过去几周,我们一直在用数十个开源项目验证这条流水线。这些初步披露包含了数百份漏洞报告,其中多个漏洞被我们串联成影响这些项目的未认证远程代码执行漏洞利用。每份报告都包含一个自包含的复现程序、对漏洞的说明(在可能的情况下,还包括通过二分查找确定漏洞引入时间),以及(在可用时)关于如何修复该漏洞的候选补丁。我们现在正将这项服务开放给更多开源项目。 在我们测试 OSS Scanner 早期版本和原始模型输出时,收到了一些反馈: - “OSS Scanner 的发现中,揭示 PostgreSQL 缺陷的比例异常之高。有几份报告附带的修复方案我们几乎可以直接使用,而快速通道访问让我们能在最新问题进入 GA 版本之前就加以解决。”——Noah Misch,PostgreSQL - “大约 18 个月前,在 Project Glasswing 之前,早期的 AI 报告糟透了。我们从 Anthropic 收到的报告,包括原始模型输出在内,质量与我们从人类那里得到的一样好,有时甚至更好。尤其是当报告附带真实漏洞利用时,对工程师来说基本就大功告成了,因为你可以立即验证。”——Anton Arapov,OpenSSL Corporation - “我们发现这些报告的信号很强:在我们收到的 74 份报告中,除两份外全部有效,其中五份成为了 CVE。由于附带了补丁,这些报告直接融入了我们现有的验证和修复流程。我们希望能收到更多。”——Todd Ouska,wolfSSL - “这些漏洞报告详尽而清晰,对 HotCRP 复杂的权限模型有深入理解,并且对漏洞的优先级排序也很合理。”——Eddie Kohler,HotCRP 为验证 OSS Scanner 的早期版本,我们请审查我们 CVD 发现结果的专家渗透测试人员,检查了扫描器在 48 个项目中发现的 97 个严重和高危漏洞。其中 85 个(88%)达到了我们 CVD 流程的标准。在其余 12 个中,11 个是真实问题,但与已知问题或本次扫描的其他发现重复,只有一个无效,即“误报”。此后我们已将许多此类发现发送给维护者,他们很少告诉我们某个高危或严重发现是无效的,这与[我们此前的开源工作](https://red.anthropic.com/2026/cvd/)一致。有些人告诉我们,严重性评级可能被夸大,或者扫描器误解了项目的威胁模型。我们无法保证扫描器完美无缺,但我们会根据维护者的反馈以及模型的改进,持续完善该系统。 ## 开始使用 符合条件的项目的核心维护者可以通过向[这个 GitHub 仓库](https://github.com/anthropics/oss-scanner)提交 PR 来注册,需遵循标准项目模板,更多指引见[我们的扩展 FAQ](https://red.anthropic.com/oss-scanner)。项目是否符合条件基于与 OSS-Fuzz 类似的一套标准:简而言之,项目应对“基础设施和用户安全具有关键影响”,我们将逐案做出决定。 _无论 OSS Scanner 是否适合你,我们新的[网络验证计划](https://www.anthropic.com/news/cyber-verification-program)都为符合条件的安全专业人员提供高级网络能力和降低拦截的分类器。[Claude for OSS](https://claude.com/contact-sales/claude-for-oss) 提供免费的 Claude Max 20x 订阅,以帮助修复漏洞和改进 OSS 项目。_ ## 相关内容 ### Claude 式科学 客座作者 Matthew Schwartz 教授描述了当他不再与 Claude 对抗、转而让 Claude 去寻找“Claude 形状”的问题时发生了什么:这类问题最适合当前一代 LLM 工具的能力。这促使他构建了 BootLoops,一个用于定量科学中精确计算的工具包,他一直在与专家合作将其应用于各个科学领域。 [阅读更多](https://www.anthropic.com/research/claude-shaped-science) ### 机器人能做什么工作? 我们构建了一个指数,用来衡量当今机器人在执行美国工作任务方面的表现。机器人已经能够完成四分之三的体力任务,但大多是在受限环境中,且仅在 0.3% 的任务上具有成本竞争力。 [阅读更多](https://www.anthropic.com/research/what-work-can-robots-do) ### 你想从 AI 中得到什么? 我们正在启动一项新研究,使用 Anthropic Interviewer 来了解你与 AI 相处的经历。 [阅读更多](https://www.anthropic.com/research/your-thoughts-on-ai) ## 正文 · 原文 _We’re launching [OSS Scanner](https://red.anthropic.com/oss-scanner), an opt-in vulnerability scanner for the open-source ecosystem informed by our experience using Claude to find vulnerabilities during [Project Glasswing](https://www.anthropic.com/glasswing). Projects that join will receive thorough, periodic security scans by our strongest models at no cost._ Language models have rapidly advanced in their ability to [discover vulnerabilities](https://www.anthropic.com/research/mythos-preview), as we have [reported extensively](https://www.anthropic.com/research/zero-days) on this blog. On [CyberGym](https://arxiv.org/abs/2506.02548), one academic vulnerability-finding benchmark, LLMs have gone from finding under 20% of vulnerabilities at the beginning of last year to finding over 85% this year. As a result, open source maintainers have gone from receiving [mostly slop](https://daniel.haxx.se/blog/2025/07/14/death-by-a-thousand-slops/) from LLMs to receiving [high-quality bug reports](https://lore.kernel.org/linux-media/CAHk-=wi4zC+Ze8e+p3tMv8TtG_80KzsZ1syL9anBtmEh5Z40vg@mail.gmail.com/). Over the last six months, we’ve used our latest models to scan for vulnerabilities in some of the world’s most important software projects. We have discovered [over 29,000 candidate vulnerabilities](https://red.anthropic.com/2026/cvd/), but have only been able to manually review and triage approximately 6,000 of these. While 6,000 vulnerabilities is significant, we remain bottlenecked on our human capacity to validate these findings. We are working on scaling our vulnerability disclosures. Meanwhile, with increasing frequency, maintainers who receive our first reports simply ask us for a bulk submission of all the unverified reports with proposed patches: to date, we have sent nearly 5,000 reports directly to maintainers after they asked to receive everything we had—even if it wasn’t validated. Since exploits can now be developed [in minutes](https://www.anthropic.com/research/n-days), projects that can find and address vulnerabilities faster can better secure their software against attackers who are racing to find and leverage these same weaknesses. We will continue to manually disclose human-verified vulnerability reports via our existing [coordinated vulnerability disclosure (CVD) process](https://www.anthropic.com/coordinated-vulnerability-disclosure), especially for projects without the resourcing to triage reports themselves. But we are now making available an optional fast-track for those who wish to receive vulnerability reports as soon as they become available. ## Our vulnerability scanner Inspired by the positive impact on the open source ecosystem of [Google's OSS-Fuzz](https://google.github.io/oss-fuzz/), a project that scans open-source software for vulnerabilities with fuzzers, we are launching [OSS Scanner](https://red.anthropic.com/oss-scanner) to find vulnerabilities in open-source code with our strongest language models. While [Claude Security](https://claude.com/product/claude-security), our general-access code scanning and patching product focuses on helping enterprises defend their systems, OSS Scanner provides security audits at no cost to open-source projects. The outputs of this opt-in vulnerability scanner will be fully model-generated, without human review or triage. This will enable faster and more frequent scanning, but means that it is possible reports will be incorrect or invalid. These reports will be generated by our strongest models (including Claude Mythos) to give open-source projects the largest defensive advantage. We have spent the last several weeks validating this pipeline with dozens of open source projects. These initial disclosures contained hundreds of bug reports, including multiple vulnerabilities that we were able to chain to unauthenticated remote code execution exploits affecting these projects. Each report contains a self-contained reproducer, explanation of the vulnerability (including a bisection to determine when the bug was introduced, where possible), and candidate patch (when available) for how the bug can be fixed. We are now making this service available to more open source projects. Some of the feedback we received as we tested early versions of OSS Scanner and our raw model outputs: - “An unusually high fraction of OSS Scanner’s findings uncovered PostgreSQL defects. Several reports came with fixes we can use nearly as-is, and fast-track access let us address the newest issues before they reached a GA release.” —Noah Misch, PostgreSQL - “Early AI reports about 18 months ago, before Project Glasswing, were appalling. The reports we received from Anthropic, raw model output included, were as good and sometimes better than what we get from people. Particularly when a report comes with a real exploit attached, that's basically job done for an engineer as you can verify it right away” —Anton Arapov, OpenSSL Corporation - “We found the signal from these reports high: of the 74 reports we received, all but two were valid, and five became CVEs. With patches attached, the reports slotted right into our existing process to verify and fix issues. We’d love more.” —Todd Ouska, wolfSSL - “The bug reports were thorough and clear, with a strong understanding of HotCRP's complex permission model and good bug prioritization.” —Eddie Kohler, HotCRP To validate an early version of OSS Scanner, we asked the expert penetration testers who review our CVD findings to check 97 critical and high-severity vulnerabilities from the scanner across 48 projects. Of these, 85 (88%) met the bar for our CVD process. Of the remaining 12, 11 were real but duplicated known issues or other findings from the scan, and only one was invalid, i.e. a “false positive.” We’ve since sent many of these findings to maintainers, who have seldom told us a high or critical finding was invalid, consistent with [our prior open-source work](https://red.anthropic.com/2026/cvd/). Some have told us severity ratings can be inflated or the scanner misunderstood the project’s threat model. We can’t guarantee the scanner will be perfect, but we’ll keep refining the system based on maintainer feedback and as models improve. ## Getting started Core maintainers of eligible projects can enroll by submitting a PR to [this GitHub repo](https://github.com/anthropics/oss-scanner) following the standard project template, with additional guidance provided in [our extended FAQ](https://red.anthropic.com/oss-scanner). Projects are eligible based on a similar set of criteria that OSS-Fuzz uses: briefly, projects should have a “critical impact on infrastructure and user security” and we will make decisions on a case-by-case basis. _Whether or not OSS Scanner is right for you, our new [Cyber Verification Program](https://www.anthropic.com/news/cyber-verification-program) makes advanced cyber capabilities and reduced blocking classifiers available to qualifying security professionals. [Claude for OSS](https://claude.com/contact-sales/claude-for-oss) provides free Claude Max 20x subscriptions to help remediate vulnerabilities and improve OSS projects._ ## Related content ### Claude-shaped science Guest author Prof. Matthew Schwartz describes what happened when he stopped fighting Claude and allowed Claude to find “Claude-shaped” problems: ones best suited to the capabilities of the current generation of LLM tools. This led him to build BootLoops, a toolkit for exact calculations in quantitative science, which he has been applying across scientific fields alongside experts. [Read more](https://www.anthropic.com/research/claude-shaped-science) ### What work can robots do? We built an index of how well today’s robots can perform US job tasks. Robots can already do three-quarters of physical tasks, mostly in limited settings, but are cost-competitive for just 0.3% of them. [Read more](https://www.anthropic.com/research/what-work-can-robots-do) ### What do you want from AI? We’re launching a new study using Anthropic Interviewer to learn from your experiences with AI. [Read more](https://www.anthropic.com/research/your-thoughts-on-ai)