```
Low Fix Rate Undermines Anthropic's Glasswing Vulnerability Push, VulnCheck Finds
Anthropic's Project Glasswing, an effort to put AI-driven vulnerability discovery to work on open-source software, has come under scrutiny after a VulnCheck analysis found that a small fraction of what the system reports ever becomes public — and an even smaller fraction actually gets fixed.
Patrick Garrity of VulnCheck, writing in a blog post summarised by LWN.net, tallied the project's output since Glasswing launched on April 7. Of 37,966 findings Anthropic reports, only 7,349 — 19.4% — have been made public. Of those public disclosures, just 217 have been marked fixed: roughly 3% of the findings the project was willing to share with the world.
The ratios, rather than the raw volume, are what make this a story worth paying attention to. Automated bug-finding tools have long been judged on how many issues they can surface; Glasswing's numbers invite a harder question: how much of that output turns into remediated code?
Detection was never the bottleneck
Vulnerability researchers and open-source stewards have argued for years that discovery is the easy half of the problem. Triaging a report, confirming it is real rather than a hallucinated or duplicate finding, reproducing it, and coordinating a responsible disclosure with a maintainer who may have no bandwidth — that is where queues stall. Glasswing's output appears to push more work onto that same constrained stage rather than relieving it.
It is fair to note that fixes naturally lag disclosures: a bug reported last month may be awaiting a release cycle, a maintainer's holiday, or a downstream consumer's review. The 217 figure is therefore a snapshot, not a verdict. But the size of the gap between 37,966 findings and 217 fixes suggests the pipeline between detection and remediation is doing most of the heavy lifting — and doing it largely by hand.
The dynamic touches a familiar sore point in the open-source community. Volunteer maintainers have spent the past decade absorbing an expanding stream of scanner output, bug-bounty reports and, increasingly, machine-generated findings, often without corresponding increases in support. An AI system that generates thousands of candidate vulnerabilities per quarter can be a genuine force multiplier for a project with a security team, and a source of fatigue for one with two people and full-time day jobs.
What developers should actually measure
For teams evaluating AI-assisted vulnerability-scanning tools, the practical takeaway is to look past detection counts in vendor materials and ask for downstream numbers: what share of findings were confirmed, disclosed and fixed, and over what timeframe. Fix rate, not report rate, is the metric that reflects real security improvement. Teams should also plan explicitly for triage capacity — the tooling shifts where the work lands, but it does not remove it — and decide in advance who owns disclosure for each incoming batch of findings.
The critique is not that Glasswing's findings are worthless — some of them correspond to real, fixable bugs — but that a high-volume, low-signal stream can impose costs on the very community the effort is meant to protect. Whether Anthropic's pipeline matures toward higher disclosure and fix rates as it accumulates data is the natural next beat in this story.
VulnCheck's full analysis is available on the company's blog, and LWN.net carried the write-up. Anthropic's own description of Project Glasswing is published on its site.
修復率偏低動搖 Anthropic Glasswing 漏洞計劃的成效，VulnCheck 研究發現
Anthropic 的 Project Glasswing 是一項將 AI 驅動的漏洞發現技術應用於開源軟件的計劃，現正受到外界審視。VulnCheck 的分析顯示，該系統所報告的漏洞中只有極小部分曾公開披露，而真正獲得修復的比例更為微小。
VulnCheck 的 Patrick Garrity 在一篇經 LWN.net 摘要的博客文章中，統計了自 Glasswing 於 4 月 7 日啟動以來的產出。在 Anthropic 報告的 37,966 項發現中，只有 7,349 項——佔 19.4%——曾予公開。在這些公開披露中，僅有 217 項被標記為已修復：約佔該計劃願意與外界分享的發現的 3%。
值得關注的是這些比率，而非原始數量。自動化漏洞搜尋工具一向以能發現多少問題作為評核標準；Glasswing 的數據卻引出一個更尖銳的問題：這些產出中，究竟有多少最終轉化為已修復的代碼？
偵測從來不是瓶頸
漏洞研究者與開源項目維護者多年來一直認為，發現漏洞只是問題中較容易的一半。分類處理報告、確認其屬實而非幻覺或重複發現、重現漏洞，以及與可能毫無餘裕的維護者協調負責任披露——這才是令工作流程停滯的環節。Glasswing 的產出看來只會把更多工作推入這個同樣受限的階段，而非為之紓緩。
值得指出的是，修復自然會滯後於披露：上月回報的漏洞可能正等待發布周期、維護者的假期，或下游用家的審核。因此 217 這個數字只是一個快照，而非最終判決。然而，37,966 項發現與 217 項修復之間的差距之大，顯示從偵測到補救的整個流程承擔了大部分艱巨工作——而且主要依靠人手完成。
這一情況觸及開源社群一個屢見的痛處。義工維護者過去十年一直承受不斷膨脹的 scanner 輸出、bug bounty 報告，以及日益增多的機器生成發現，卻往往沒有獲得相應的支援增加。一個每季生成數千項候選漏洞的 AI 系統，對擁有安全團隊的項目而言可以是真正的戰力倍增器，但對只有兩人兼且有全職日間工作的人來說，卻是疲勞的來源。
開發者實際上應衡量甚麼
對於評估 AI 輔助漏洞掃描工具的團隊，務實的建議是：不要只看供應商材料中的偵測數字，而要追問下游數據——有多少比例的發現獲確認、披露及修復，以及所需時間有多長。反映真實安全改善的指標是修復率（fix rate），而非報告率（report rate）。團隊亦應明確規劃分類處理（triage）能力——工具只是轉移了工作落點，並沒有消除工作——並事先決定每批新發現由誰負責披露。
批評並非指 Glasswing 的發現毫無價值——其中部分確實對應真實、可修復的 bug——而是指高流量、低訊號的資訊流可能對計劃理應保護的社群本身構成代價。Anthropic 的流程會否隨數據累積而逐步提升披露率與修復率，將是這則報導的自然下一章。
VulnCheck 的完整分析可在該公司的博客查閱，LWN.net 則刊載了相關報導。Anthropic 對 Project Glasswing 的自我介紹亦已在其網站發布。