Back to home

Compare

Comparing: Seven Months Later, AI Security Still Isn’t Helping Ordinary Maintainers & 七个月后,AI 安全项目为何仍没帮到普通维护者

AEN
Glasswing开源维护者小团队安全·

Seven Months Later, AI Security Still Isn’t Helping Ordinary Maintainers

When I first saw “automatically stop dependency attacks,” I thought of my own small projects

Last Tuesday night, I was still working on a client website in my Hangzhou rental apartment when a risk alert suddenly appeared in the admin panel. My first reaction was to upgrade, but there were too many components, and I got stuck on which one actually needed attention. When the news said Glasswing could automatically handle this kind of supply-chain risk, I was tempted too: it sounded like someone finally had maintainers’ backs.

What it actually solves, and who is using it

Glasswing is Anthropic’s software supply-chain security project. Its focus is not writing code for people, but identifying attacks that could come through third-party dependencies and helping maintainers take action. A public update says that, after seven months in operation, its results were weaker than expected, showing that there is still a gap between “automatically discovering risks” and “actually reducing incidents.” Security teams such as VulnCheck have continued tracking the results. It is a reminder that security tools have to fit into real maintenance workflows, not just demos.

What it costs to replicate today, and the first step

If I wanted this capability directly, the more realistic option today would be applying for beta access rather than building my own setup from news reports. Based on currently available public plans, the expected cost is 0 yuan, with about 2 hours of preparation. The technical barrier is being able to understand code repositories, dependency updates, and test workflows. The first step is to open Glasswing’s official project page, apply or join the waitlist, and export my dependency inventory at the same time. I wouldn’t change production just to chase a trend. I made that mistake, and after we focused only on upgrading, my colleagues spent three extra hours on regression testing that day.

How I would choose at different stages

If I were just starting out, mainly building websites or content services, I would keep automatic updates and assign one person to manually review them; there’s no urgency to try Glasswing now. With 1–2 clients, I would organize the dependency inventory, backups, and rollback steps before assessing whether to apply for beta testing. While scaling up, I would add supply-chain security to the delivery checklist and ask suppliers to explain how they handled risk discoveries. Not everyone needs this tool. If I weren’t maintaining any code at all, asking the hosting provider for a clear answer would probably save more time.

BZH
Glasswing开源维护者小团队安全·

七个月后,AI 安全项目为何仍没帮到普通维护者

我第一次看到“自动防住依赖攻击”,先想到的是自己的小项目

上周二晚上,我还在杭州的出租屋里改客户网站,后台突然弹出依赖库有风险。我第一反应是升级,可组件太多,真正该动哪一个,我也卡过。新闻里说 Glasswing 能自动处理这类供应链风险,我也很心动:听起来,终于有人替维护者兜底了。

它解决的到底是什么,谁已经在用

Glasswing 是 Anthropic 面向软件供应链安全的项目,重点不是替人写代码,而是识别第三方依赖可能带来的攻击,再协助维护者采取行动。公开更新提到,项目运行七个月后,发现的成果与预期相比偏弱,说明“自动化发现风险”和“真正减少事故”之间仍有距离。VulnCheck 等安全团队持续跟踪相关结果;这提醒我们,安全工具还得放进真实维护流程里,而不能只看演示。

今天复刻的成本,以及第一步

如果你想直接获得这项能力,现阶段更现实的是申请测试资格,而不是照着新闻自己搭一套。费用按公开可用计划计算:预计 0 元;准备时间约 2 小时;技术门槛是能看懂代码仓库、依赖更新和测试流程。第一步:打开 Glasswing 官方项目页面,点申请或加入候补名单,同时先把自己的依赖清单导出来。不要为追热点改生产环境,这步我搞错过,结果只顾着升级,反而让同事当天多花了三小时回归测试。

如果你处在不同阶段,我会这样选

如果你刚起步、主要做网站或内容服务,我会先保留自动更新,并留一个人工复查;现在不试也没事。如果已经有 1—2 个客户,我会整理依赖清单、备份和回滚步骤,再评估是否申请测试。若你正在扩规模,我会把供应链安全纳入交付检查表,并要求供应商说明发现风险后的处置记录。这工具不是所有人都需要,尤其你完全不维护代码时,先找托管服务商问清楚更省时间。