
Anthropic 推出 OSS Scanner:面向开源项目的免费 AI 漏洞扫描器
Anthropic 于 10 月 8 日正式推出 OSS Scanner,这是一款免费、可自愿加入的漏洞扫描服务,专为关键开源项目设计。它由 Anthropic 的 Frontier Red Team 与 Safeguards Research 团队维护,借鉴了 Google OSS-Fuzz 的思路,但用大语言模型替代了模糊测试器(fuzzer)。
核心维护者可通过在 anthropics/oss-scanner 仓库提交 Pull Request 完成注册,提供 repo 地址、主要联系人以及 Dockerfile 路径。扫描在与互联网隔离的虚拟机中运行,使用 Anthropic 最强模型(含 Claude Mythos)对代码库进行定期审计,报告直接发送给维护者,无需人工审核。
上线 11 天内,截至 10 月 19 日,OSS Scanner 已从约 650 份申请中接纳 114 个项目;累计报告 77 个严重级(critical)与 803 个高危级(high)由模型评级的漏洞。在 97 项经审计的严重与高风险发现中,85 项(88%)符合 Anthropic 的披露标准,仅 1 项为误报;以 wolfSSL 为例,74 份报告中有 72 份有效。但也存在 11 项与已知问题重复、个别维护者夸大严重性的情况。
过去六个月里,Anthropic 的模型已发现超过 29,000 个潜在漏洞,人工审查员仅对其中的约 6,000 个进行了分类,另有近 5,000 份报告直接转交维护者。OSS Scanner 将这种临时请求转化为一项规范的自愿计划,标志着 AI 驱动的开源供应链安全审查正从"锦上添花"走向行业预期标准。
本文来源:https://ncijnetwork.com/anthropic-launches-oss-scanner-a-free-ai-vulnerability-scanner
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
Anthropic Launches OSS Scanner: A Free AI Vulnerability Scanner for Open Source Projects
On October 8, Anthropic officially launched OSS Scanner, a free and opt-in vulnerability scanning service designed for critical open source projects. It is maintained by Anthropic’s Frontier Red Team and Safeguards Research teams, draws inspiration from Google OSS-Fuzz, but replaces the fuzzer with large language models.
Core maintainers can register by submitting a Pull Request to the anthropics/oss-scanner repository, providing the repo address, a primary contact, and the Dockerfile path. Scans run in network-isolated virtual machines, using Anthropic’s most powerful models (including Claude Mythos) to periodically audit codebases, with reports sent directly to maintainers without manual review.
Within 11 days of launch, as of October 19, OSS Scanner accepted 114 projects out of approximately 650 applications; it has reported a total of 77 critical and 803 high model-rated vulnerabilities. Among 97 audited critical and high-risk findings, 85 (88%) met Anthropic’s disclosure criteria, with only 1 false positive; for example, wolfSSL had 72 valid reports out of 74. However, there were also 11 duplicates of known issues and a few cases where maintainers overstated severity.
Over the past six months, Anthropic’s models have identified more than 29,000 potential vulnerabilities, of which human reviewers classified only about 6,000, and nearly 5,000 reports were passed directly to maintainers. OSS Scanner turns this ad hoc request process into a structured, voluntary program, signaling that AI-driven security review of the open source supply chain is moving from a nice-to-have to an industry-expected standard.