尽管 Anthropic 的 Project Glasswing 项目在为期七个月的跟踪期间生成了大量安全发现,但其实际成效令开源维护者感到失望 1。数据显示,该项目仅促成了全年新增 CVE 总数的 0.5%,且绝大多数发现并未得到修复或证实为有效威胁 1。具体而言,截至统计时,Anthropic 追踪到 442 个归因于其发现的 CVE,约占今年发布的 93,450 个 CVE 总量的微小比例 1。在这 442 个 CVE 中,只有两个(CVE-2026-61500 和 CVE-2026-26980)被确认在野外遭到利用 1。
项目的内部账本进一步揭示了高误报率与低修复率的困境 1。Anthropic 账本显示共有 37,966 项发现,其中仅有 7,349 项公开,而标记为“已修复”的数量仅为 217 项 1。此外,有 1,654 项(占比 4.4%)被判定为重复,另有 419 项被撤回 1。自九月以来,新增了近 12,000 项发现,但修复数量仅增加了 15 项 1。这种巨大的反差使得 Anthropic 声称的真实阳性率从启动时的 90.8% 上升至 94.2% 的说法引发质疑 1。
随着十月八日 OSS Scanner 工具的推出,验证负担被直接转移给了开源项目维护者 1。该工具允许自助扫描,但明确说明报告由模型生成且未经人工审核 1。以 curl 项目为例,维护者 Daniel Stenberg 收到了 Anthropic 提交的 45 份报告 1。在目前已公开的 34 份报告中,有 19 份被标记为非漏洞,6 份确认为低严重性,剩余 9 份待评估,由此计算出的误报率达 55.9% 1。
A seven-month evaluation of Anthropic’s Project Glasswing reveals that despite generating a massive volume of reports, the initiative has resulted in negligible real-world security improvements for open-source software. As of the latest statistics, Anthropic is credited with 442 Common Vulnerabilities and Exposures (CVEs), representing just 0.5% of the 93,450 CVEs published globally this year 1. The impact on active threats remains minimal, with only two of these identified vulnerabilities—CVE-2026-61500 and CVE-2026-26980—confirmed as exploited in the wild 1. While Anthropic claims its "true positive rate" has improved from 90.8% at launch to 94.2%, independent analysis of the project's ledger suggests significant discrepancies between reported findings and actual remediations 1.
The data indicates that the sheer quantity of discoveries has not translated into fixes, creating an administrative burden for maintainers rather than enhancing security. Out of 37,966 total findings recorded by Anthropic, only 7,349 were made public, and merely 217 were marked as fixed 1. A notable 1,654 findings (4.4%) were labeled as duplicates, while 419 were withdrawn entirely 1. This stagnation persists even as new discoveries accumulate; since September, nearly 12,000 additional findings have been added to the ledger, yet the number of confirmed fixes increased by only 15 1.
The rollout of Anthropic’s OSS Scanner tool in October has further exacerbated tensions by shifting verification responsibilities directly onto open-source projects without human review. Although the scanner allows projects to self-audit, Anthropic explicitly notes that all generated reports are unverified by humans 1. Daniel Stenberg, maintainer of the curl project, received 45 such reports, highlighting the high false-positive rate inherent in the system 1. Of the 34 reports currently public regarding curl, 19 were classified as non-vulnerabilities and six as low-severity issues, leaving nine under evaluation; this results in a 55.9% misidentification rate for the initial batch 1.
评论
还没有评论,欢迎留下第一条。