最新资讯
dev.to 在信任 AI 生成的测试用例前,如何逐条复核 unite.ai Franklin Templeton 构建 AI 研究工作流以揭示风险 techxplore.com AI 助力寻找更优的钛合金 3D 打印工艺参数 engadget.com Perplexity 推出“teammates”日常工作任务 AI agent computerworld.com 别提数据中心:OpenAI 改称“compute campuses” theguardian.com 俄美试图阻止就自主武器制定具有约束力的规则 bloomberg.com Nvidia 加强东南亚芯片追踪以打击走私 thestack.technology 模型路由需求升温,Agent Router 推出开源替代方案 www.youtube.com Google DeepMind 的 Philipp Schmid:AI 同事需要具备的七个要素 unite.ai Cashfree 为 AI Agent 的托管结账加入 WebMCP 工具 dev.to 我们在 DolphinBench 上测试记忆检索:纯关键词搜索暴露了自身回归 dev.to 别再把 Agent 写成 while 循环,把它建模为状态机
多元动态
登录 注册
智能体 dev.to · @Unite and Create For Life

GitHub pull_request_target 变更在千星仓库中破坏了什么

原文 我的收藏

作者用自建检查工具 prt-check 扫描 GitHub 上 1000 个星标最多的仓库,发现 269 个(26.9%)至少有一个工作流使用 pull_request_target,涉及 540 个工作流文件。文中列出两项带日期的 GitHub 变更:自 2026-07-20 起 actions/checkout 默认拒绝在 pull_request_target 与 workflow_run 中检出 fork PR 代码,除非显式选择加入;自 2026-11-02 起 GitHub 将默认封禁未配置 Actions 策略的公开仓库使用该触发器。报告仅公布聚合数字,不点名具体仓库。

AI推荐理由 给出可核验的具体数据:扫描 GitHub 千星仓库发现 269/1000(26.9%)至少有一个工作流使用 pull_request_target,共 540 个工作流文件;并列出两项有日期的 GitHub 变更(2026-07-20 checkout 限制、2026-11-02 默认封禁)。工具由作者自建,报告仅给聚合数字、不点名仓库,方法透明度尚可。

66

正文

GitHub is changing how pull_request_target works this year. Two dated changes affect every public repository that uses the trigger, so I built a small free checker, prt-check, with my AI engineering system (HAL), and ran it over the 1,000 most-starred repositories on GitHub. Check your own repository in one step: - uses: actions/checkout@v4 - uses: UniteAndCreateForLife/prt-check@v1 Enter fullscreen mode Exit fullscreen mode Or run python prt_check.py locally (no dependencies). The two changes: since 2026-07-20, actions/checkout refuses to check out fork pull request code in pull_request_target and workflow_run workflows unless you opt in (changelog). From 2026-11-02, GitHub blocks pull_request_target on public repositories that have no Actions policy allowing it (changelog). Scanned 2026-09-26, 37 days before GitHub's default block on 2026-11-02, with prt-check. Aggregate numbers only: this report names no repository. The numbers 269 of 1,000 (26.9%) run at least one workflow on pull_request_target (540 workflow files). Unless their maintainers allow the trigger in an Actions policy, those workflows stop running on 2026-11-02. 9 (0.9%) check out fork pull request code in a privileged workflow, with an actions/checkout that has the new guard and no condition that keeps forks out. Since 2026-07-20 the guard refuses those checkouts, so the steps fail for pull requests from forks. Label-gated ones fail when the label is added. 3 check out fork code in a privileged workflow with actions/checkout pinned to a version or commit from before the guard. The guard doesn't apply to them, so the fork's code is checked out where the workflow's token and secrets are. They are the most urgent to fix, and this report does not name them. 4 opted in with allow-unsafe-pr-checkout: true: a deliberate decision, safe only if no later step executes the fork's files. 9 fetch pull request code with git fetch ...pull/... or gh pr checkout in a privileged workflow. That path bypasses the new guard in actions/checkout. 8 run an AI or review action on pull_request_target. Those reviewers go quiet on 2026-11-02 unless the trigger is allowed. What runs on pull_request_target today The actions used most often in these workflows, by the number of repositories that use them there: Action Repositories actions/github-script 125 actions/labeler 50 actions/create-github-app-token 25 actions/setup-node 22 actions/setup-python 16 amannn/action-semantic-pull-request 13 eps1lon/actions-label-merge-conflict 12 actions/upload-artifact 12 contributor-assistant/github-action 10 step-security/harden-runner 6 dorny/paths-filter 6 actions/download-artifact 6 AI and review actions on the trigger: Action Repositories anthropics/claude-code-action 5 presubmit/ai-reviewer 1 anthropics/claude-code-base-action 1 openai/codex-action 1 Method Selection: the 1,000 most-starred public, non-fork, non-archived repositories, from GitHub's repository search on 2026-09-26. 809 of them have workflow files (9,328 files in total). 0 could not be read. Only .github/workflows/*.yml and *.yaml on the default branch were read, through the GitHub API and raw.githubusercontent.com. No workflow was run, and nothing was written anywhere. Each file went through prt_check.py, which reads YAML line by line. README lists the codes, and the tests show the layouts it understands. Every repository behind the second and third numbers was reviewed by hand, and the checker was corrected wherever the review found it wrong: job conditions, Dependabot-only jobs, workflow_run upstreams and branch filters, and commit pins resolved to exact actions/checkout releases. Actions policies (repository, organization or enterprise) are not visible from outside, so some of these repositories may already allow pull_request_target and keep running it after 2026-11-02. The numbers count workflows that depend on the trigger, not workflows that will certainly stop. What to do Run prt-check on your repository (Action or one command), then follow the three fixes in the README: switch to pull_request, split into pull_request + workflow_run, or keep the trigger deliberately with an Actions policy. Source code, tests and the full method: https://github.com/UniteAndCreateForLife/prt-check. If you hit a false positive, open an issue and I'll fix it.

来源报道 · 1 篇 · 以下链接均前往原始媒体

Unite and Create For Life GitHub pull_request_target 变更在千星仓库中破坏了什么 09-27 · 原文发布
← 返回精选动态 原文

更多内容