< h 2 > Scroll ing through the vulnerability advisory last Friday night , my palms were sweating </ h 2 >< p > Scroll ing through this vulnerability advisory last Friday night , my palms were sweating —all my client projects are on GitHub . If you use GitHub to store code or run automated deployments , this vulnerability called CVE - 202 6 - 38 5 4 is worth 10 minutes of our time . Simply put , someone could remotely execute arbitrary code on your server through GitHub 's automation feature called Actions . It sounds scary , but GitHub has already fixed it ; what we need to do is self -check our settings .</ p >< h 2 > What this bug is — Aj ie almost got hit </ h 2 >< p > Dis covered by the W iz security team , this vulnerability exploits GitHub Actions ' permission mechanism . My friend Aj ie , who runs an independent design studio , was modifying a client 's project code in a cafe in Hang zhou last Wednesday when he found an unidentified script running in his Actions log — fortunately , it was just a test repo ; it would have been a disaster if it were the production environment . I 've made similar mistakes before : for convenience , I gave write permissions to Actions for all repositories , which was like leaving the front door wide open for anyone . This vulnerability doesn 't discriminate ; as long as we use Actions automation , there 's a risk .</ p >< h 2 >Your self -check cost today </ h 2 >< p >$ 0 + 10 minutes + Technical barrier : Just be able to log into the GitHub backend and check settings . First step : Log into GitHub , click into Settings of any repo , find Actions → General on the left , and see if " Workflow permissions " is set to " Read " or " Read and write " . If it 's the latter , and we don 't have a specific scenario requiring write permissions , changing it back to Read is fine . This tool isn 't needed for everyone —if you 've never turned on Actions , it 's fine not to check now .</ p >< h 2 > Advice by stage </ h 2 >< p > Just starting out : If we only use GitHub to store code and haven 't touched Actions automation , the impact is minimal . Just glance at the permissions next time you log in , no rush . 1 - 2 clients : If we use Actions to auto -de ploy websites , I 'd suggest checking the Actions permissions for every repo today , and whether the third -party actions referenced in the workflow files are from trusted sources . Scaling up : If multiple team members have admin privileges , I recommend immediately auditing all repo permissions , enabling branch protection , and restricting who can modify workflow files . I got stuck here before too , spending a whole afternoon sorting it out , but it was worth it .</ p >
GitHubSecurity Vulner abilitySol op rene urPersonal IPSmall Team··3 min read·chatopc.com·via www.wiz.io·
Your repo could be remotely hij acked — GitHub bug fixed , 10 -min check
相关推荐
同分类:ai_news
Qwen 2.5 7BApostate
三种工具都能拆掉模型“安全阀”,这说明开源大模型的护栏并不牢靠
一组针对 Qwen 2.5 7B 的测试显示,3 个不同工具都能把模型对有害请求的拒绝率几乎清零,最好的一种甚至做到 100% 服从。这不只是“越狱工具”又多了一个,而是再次提醒我们:开源模型的安全训练并没有外界想得那么稳固。
6月3日·www.reddit.com
DolphinGemmaLocalLLaMA
DolphinGemma 迟迟未发,开源模型热度高但交付正在变得更难
一条 Reddit 追问贴本身不是新闻,但它点出一个更值得关心的事实:开源大模型圈里,“先预告、后跳票”正在变常见。DolphinGemma 至今未见正式发布,说明模型竞争已不只比效果,也开始比团队的交付能力与持续维护能力。
6月2日·www.reddit.com
MiniMaxMiniMax M3
MiniMax M3 被指几乎不设政治审查,这对中国大模型是个危险信号
一则来自开发者社区的测试称,MiniMax M3 在政治敏感话题上明显比同类中国模型更少设限。消息本身还未经官方确认,但它值得关心,因为这不是单纯的“能不能答”,而是中国大模型公司在出海、合规和产品边界之间如何取舍的问题。
6月2日·www.reddit.com
googlegemini
Google AI 眼镜接近可卖点
Google 展示 Android XR 原型眼镜,把 Gemini 翻译、导航直接叠到视野里。真正值得看的是:模型分发入口开始从手机屏幕转向脸上设备,API、agent 与 context 的竞争边界会变。
5月22日·techcrunch.com
googlegemini
Google 把 Gemini 变成入口税
Google I/O 2026 表面是 Gemini、Search 和 smart glasses 的产品更新,实质是把模型、分发和 agent runtime 重新绑成一个平台包,试图把 AI 的价值捕获点从 model quality 转回默认入口。
5月19日·www.wired.com
anthropicopenai
Karpathy 去 Anthropic 的信号
2026 年 5 月,Andrej Karpathy 加入 Anthropic pre-training team。表面是顶级研究者流动,实质是 frontier lab 竞争重新回到 pre-training、data 与 research taste,而不只是 product packaging
5月19日·techcrunch.com