AI 公司 Anthropic 推出 OSS Scanner，这是一项免费、可选择加入的服务，用于发现开源项目中的漏洞。已注册的维护者会收到由该公司最强模型（包括 Claude Mythos）生成的定期安全报告。报告未经人工审核直接发送，项目维护者可自行验证扫描结果，并决定采纳哪些修复方案。
在 10 月 8 日的公告中，Anthropic 的 Frontier Red Team 解释说，人工验证已成为瓶颈。在六个月的时间里，其模型识别了超过 29000 个潜在漏洞，但研究人员仅人工审查并分类了约 6000 个。希望提前获取报告的维护者，在提出申请后，已经收到了近 5000 份未经核验的报告和相应的修复补丁。
每份扫描报告都包含一个独立的复现程序和对疑似漏洞的说明。在可能的情况下，它还会通过二分法定位引入缺陷的代码变更，并提供候选补丁。该项目是 Glasswing 计划的延续，InfoQ 曾在 4 月份对此进行了报道，当时 Anthropic 仅向合作机构开放 Mythos 预览版访问权限。
PostgreSQL 维护者 Noah Misch 表示，快速通道访问帮助项目在正式版发布前处理了相关问题：
“有多份报告附带的修复方案几乎可以直接使用”
该公司通过一个早期版本的扫描器进行测试，邀请渗透测试人员评估了 48 个项目中的 97 个严重高危发现，其中有 85 个符合协同漏洞披露标准，另外 11 个是真实存在的漏洞，但与已知问题或其他扫描结果重复，还有 1 个是无效的。该结论仅针对本次抽样样本，不代表这项服务所有报告的准确率。
不过 Anthropic 也承认，部分维护者反馈存在漏洞严重等级虚高、对项目威胁模型理解偏差的情况。但 OpenSSL 公司的 Anton Arapov 评价，原始报告“质量不错，有时甚至高于人工提交的报告”，尤其是当有可利用的漏洞的复现代码时，验证工作会简单很多。
wolfSSL 的 Todd Ouska 报告称，候选补丁符合该项目现有的验证和修复流程：
“我们收到的 74 份报告里，仅有 2 份无效，其中 5 个被分配了 CVE 编号。”
公告之外还有一个经维护者确认的案例。curl 作者 Daniel Stenberg 将一个 HTTP/2 上传 use-after-free 漏洞归功于 Anthropic 的 Luna Tong 使用 OSS Scanner 发现。该漏洞已在 curl 8.23.0 中修复，编号 CVE-2026-92392，在特定条件下可能允许远程代码执行。Stenberg 将其定级为高危而非严重，因为成功利用该漏洞会很困难，更可能的结果是导致程序崩溃。
Anthropic 表示该项目灵感源自谷歌的 OSS-Fuzz，但 OSS Scanner 依靠大语言模型来分析代码。另一项可对比的成果是谷歌 DeepMind 的 CodeMender，InfoQ 曾在 2025 年 10 月报道过。在 CodeMender 的发布公告中，研究员 Raluca Ada Popa 和 Four Flynn 介绍了基于程序分析和测试进行的自动补丁验证。他们在提交上游代码前保留人工审核环节，这与 OSS Scanner 直接推送未审核结果的模式不同。
发现漏洞之后的工作也成为社区评论的焦点。Varun Badhwar 在 LinkedIn 上写道：“如果你的版本还停留在三个大版本之前，上游补丁也帮不了你什么。”Badhwar 是应用安全厂商 Endor Labs 的联合创始人兼 CEO，他认为企业需要确定易受攻击的代码是否可达，以及是否可以在不破坏应用程序的情况下采用修复方案。
Dimitrios Thomas 在 LinkedIn 上以个人身份写道，评估指标应包含“验证准确性、维护者工作量以及从发现到部署修复的时间”，评价维度不能只看漏洞上报数量。
扫描器的 FAQ 页面描述了智能体会在报告通过电子邮件发送之前复查错误、提出补丁并分析根本原因。这种自动化检查与人工审核是分开的。后续扫描会查找新引入的问题和之前遗漏的问题，扫描频率取决于项目使用规模、服务负载等因素。
核心维护者通过提交 GitHub PR 完成注册，PR 中包含项目配置以及用于构建环境、安装依赖的 Dockerfile。配置完成后，扫描过程会在断网环境下运行。还可以选填威胁模型文件，划定扫描范围与严重等级预期。资格审核是逐案进行的，优先考虑那些对基础设施和用户安全至关重要、且已有能力处理已验证高危报告的成熟项目。
维护者可以借助威胁模型文件来识别不受信任的输入、排除部分组件，解释他们如何判断严重性，也可以指定报告格式和补丁期望。Anthropic 会人工审核注册申请，确认申请人是项目核心维护者，并且报告能够送达项目的授权安全联系人。
项目可以暂停接收报告或申请退出。未经核验的漏洞结果没有强制 90 天披露时限，但如果后续人工核验确认漏洞属实，会转入独立的协同披露流程，90 天期限从通知维护者核验结果时开始计算。该公司依旧提供人工审核后的漏洞披露服务，特别是针对那些没有资源来分类原始发现的项目。
查看英文原文：https://www.infoq.com/news/2026/10/anthropic-oss-security-scanner/
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。
AI Anthropic launched OSS Scanner, a free, optional service to detect gaps in open source projects. Registered maintainers receive periodic security reports from the company ' s strongest models, including Claude Mythos. The reports were sent directly without manual clearance and the project maintainer was free to validate the results of the scan and decide which rehabilitation options would be adopted.
In its announcement of 8 October, the Anthropic Frontier Red Team explained that manual validation had become a bottleneck. Over a period of six months, the model identified more than 29,000 potential loopholes, but only about 600 were manually reviewed and classified by researchers. Maintainers who wish to obtain reports in advance have received nearly 5,000 unverified reports and corresponding patches upon application.
Each scanned report contains a separate reproduction procedure and a description of suspected loopholes. Where possible, it also introduces faulty code changes through dichotomy and provides candidate patches. The project was a continuation of the Glasswing project, which InfoQ reported on in April, when Anthropic was only open to cooperating agencies for Mythos preview access to permission.
According to PostgreSQL maintainer Noah Misch, the fast-track access help project dealt with issues before the official version of publish:
“Many reports of accompanying rehabilitation programmes are available almost directly”
Through an earlier version of the scanner, the company invited infiltration testers to assess 97 high-risk findings in 48 projects, 85 of which met the criteria for the disclosure of synergy gaps and 11 were real gaps, but duplicated known problems or other scanning results, and one was invalid. This conclusion is only for this sample sample and does not represent the accuracy rate of all reports from this service.
However, Anthropic also acknowledged that some of the defenders responded with serious gaps and high levels of error in understanding the project threat model. However, Anton Arapov of OpenSSL assessed that the original report was “of good quality and sometimes even higher than that submitted manually”, especially when there was a modern code of available loopholes.
Todd Ouska of WolfSSL reports that candidate patches match the existing validation and restoration process of the project:
“of the 74 reports we received, only 2 were invalid, 5 of which were assigned CVE numbers.”
The bulletin was supplemented by a case confirmed by the defender. Curl author Daniel Stenberg attributed the use-after-free loophole of HTTP/2 to the Anna Tong found using OSS Scanner in Anthropic. The loophole has been repaired in Curl 8.23.0 with CVE-2026-92392, which may allow remote code execution under certain conditions. Stenberg classified it as high-risk rather than serious, as it would be difficult to successfully exploit the loophole, and more likely the result would be a collapse of the process.
Anthropic indicates that the project is inspired by Google's OSS-Fuzz, but OSS Scanner uses a large-language model to analyze the code. Another contrastable result is Google DeepMind's CodeMender, InfoQ reported in October 2025. In CodeMender ' s publish bulletin, Fellows Raluca Ada Popa and Four Flynn presented the automated patch validation based on program analysis and testing. They retain manual clearances prior to submitting upstream codes, which are different from the models in which OSS Scanner directly sends unaudited results.
Efforts to identify gaps have also been the focus of community comments. Varun Badhwar wrote on LinkedIn: “If your version is still in front of three large versions, the upstream patch will not help you.” Badhwar is the co-founder and CEO of the application safety manufacturer Endor Labs, who believes that the enterprise needs to determine whether the code is vulnerable to attack and whether it can be repaired without disrupting the application.
Dimitrios Thomas wrote on LinkedIn in his personal capacity that the evaluation indicators should include “validation of accuracy, maintenance workload and time between detection and deployment of recovery”, and that the evaluation dimension should not be limited to the number of gaps reported.
The FAQ page of the scanner describes AI agent as reviewing errors, proposing patches and analysing root causes before the report is sent by e-mail. Such automated checks are separate from manual audits. Follow-up scans will identify newly introduced and previously omitted issues, the frequency of which will depend on factors such as the size of the project, the service load, etc.
Core maintainer completes the registration by submitting GitHub PR, which contains project configurations and Dockerfiles for environmental construction and installation. When the configuration is complete, the scanning process will run in an off-grid environment. It is also possible to fill in the threat model document and define the scope of the scan and the expected level of severity. Qualifications are carried out on a case-by-case basis, giving priority to mature projects that are critical to infrastructure and user security and have the capacity to handle proven high-risk reports.
Maintainers can use threat model files to identify untrustworthy inputs, exclude parts of components, explain how they judge severity, or specify reporting formats and patch expectations. Anthropic will manually review the registration application, confirm that the applicant is the core maintainer of the project and report being able to reach the authorized security contact for the project.
Projects may suspend the receipt of reports or request withdrawal. Unverified leakage results do not impose a 90-day disclosure time limit, but if a subsequent manual verification confirms that the loophole is true, the 90-day period is transferred to an independent and synergistic disclosure process, starting with the notification of the maintainer ' s verification results. The company continued to provide manual audit-based leak disclosure services, particularly for projects that did not have the resources to classify the original discovery.
See the original English language version: https://www.infoq.com/news/2026/10/antropic-oss-security-scanner/
Auto-fetched and translated by AI Atlas. All rights belong to the original author.