供应链风险审核员

人工智能 社区

技能解读

当前原文的自动摘译,尚未人工审校。

为什么是脚本进行测量,而不是您

本报告中的每一个数字都是关于其他人项目的声明,在围绕脚本重建这项技能之前,手工收集的数据是错误的:GitHub 贡献者统计称,有超过五人维护着 lodash,而 npm 的 ACL 说是 1,而 gh 每周移动 1.64 亿次的软件包的下载量为零。不要通过gh、网络搜索或

工作流程

确认目标目录具有清单:package.json、pyproject.toml、requirements*.txt或go.mod。如果不存在,请说出来并停止 - 不要审核该收集器未手动解析的生态系统。锁定文件读取精确版本和传递扫描:package-lock.json/npm-shrinkwrap.json、uv.lock和 go 1.17+ go.mod。 yarn.lock、pnpm-lock.yaml和poetry.lock未读取 —

您添加的内容的样式

按照安全报告阅读的方式编写附加的散文,并将相同的寄存器应用于报告附录和最终答复 - 答复将逐字粘贴到票证和报告中。陈述发现、其背后的数据以及行动。非人称和声明性:没有第一人称或第二人称(“我运行了收集器”,“你应该升级”),没有缩写,没有感叹号。

阅读报告

无法评估并不是风险。 PyPI 不发布维护者 ACL,Go 也没有注册表;这些行说的是不可知的事情,而不是错误的事情。承保范围表限制了每项索赔。 “无建议”的意思是“评估的内容中没有任何建议”——在重复干净的判决之前检查评估的计数。逐字引用数字。请勿重新推导、四舍五入或修饰报告的内容

拒绝的理由

“gh 可以比收集器更快地为我提供维护者计数。”衡量错误——回购协议贡献者和注册发布权是不同的人群。 “没有任何发现,所以依赖关系是安全的。”阅读覆盖率表;在 PyPI 和 Go 上,一半的标准在结构上是无法评估的。 “无法评估的行只会让读者感到困惑;我会放弃它们。”他们是

流狐档案 运行与权限由流狐静态检查推断,不代表作者声明或实际测试结果。
领域
人工智能
作者声明的智能体
未找到明确声明;不据此推断已兼容或已测试
静态检查
88 / 100 · 启发式扫描,不代表运行安全
运行环境证据
未执行;当前没有隔离运行测试证据
作者 / 版本 / 许可证
@trailofbits · 未声明许可证
上下文用量评级
低消耗
接入复杂程度
需简单配置
是否需要外部接口密钥
不需要
兼容的系统
macOS · Linux · Windows
运行环境要求
未检出明确要求
检测到的文件与系统行为
  • 只读
  • 允许写入 / 修改
  • 命令执行
网络行为
允许外网请求
安装命令数
无(仅作为资料)

需要注意: 未声明工具权限范围;实际权限由运行环境决定。

作者说明文件(原文) SKILL.md

# Supply Chain Risk Auditor

Generates a supply-chain risk report for a project's direct dependencies (npm, PyPI,
Go), plus an advisory sweep of everything its lockfile resolves. Two deterministic
scripts do the measuring; your job is the judgment they refuse to automate.

## Why the scripts do the measuring, not you

Every figure in this report is a claim about somebody else's project, and hand-collected
figures were measured wrong before this skill was rebuilt around scripts: GitHub
contributor counts said five-plus people maintain `lodash` where npm's ACL says one, and
`gh` saw zero downloads for a package that moves 164 million a week. Do not estimate
maintainer counts, downloads, staleness, or CVE history from `gh`, web search, or
memory — run the collector, and quote what it measured.

The scripts enforce two rules worth knowing before you read their output:

- **Unavailable data is never evidence of risk.** Every criterion resolves to
  assessed-clean, assessed-flagged, or unassessable-with-a-reason.
- **An absent measurement is never a clean verdict.** A run that measured nothing exits
  non-zero instead of printing a report that finds nothing.

## Workflow

1. Confirm the target directory has manifests: `package.json`, `pyproject.toml`,
   `requirements*.txt`, or `go.mod`. If none exist, say so and stop — do not audit an
   ecosystem this collector does not parse by hand. Lockfiles read for exact versions
   and the transitive sweep: `package-lock.json`/`npm-shrinkwrap.json`, `uv.lock`, and
   a go 1.17+ `go.mod`. `yarn.lock`, `pnpm-lock.yaml`, and `poetry.lock` are not read —
   the report says so when they are present, and versions fall back to pins or the
   latest release.
2. Check `gh auth status`. Unauthenticated GitHub allows 60 requests/hour against 5,000,
   and the collector makes several per dependency; expect repository criteria to come
   back unassessable without it. Say so rather than fixing it silently.
3. Collect, then render. Put outputs somewhere outside the audited repository unless
   asked otherwise:

   ```sh
   uv run {baseDir}/scripts/collect.py <project-dir> --json <out-dir>/findings.json
   uv run {baseDir}/scripts/render.py <out-dir>/findings.json --out <out-dir>/report.md
   ```

   Expect a few minutes for ~50 dependencies — several HTTP requests per dependency,
   more with many Go modules, and slower without authenticated `gh`. If `collect.py`
   exits non-zero, it is refusing to report — relay its message verbatim instead of
   retrying or working around it.
4. Read `report.md` and `findings.json`. The report is the deliverable; the JSON carries
   the datum behind every verdict when you need to cite one.
5. Add what the collector cannot, clearly separated from what it measured:
   - A short narrative for this reader: what to act on first, and why.
   - Upgrade paths for advisory findings — check whether the fix is a patch or a major
     version away.
   - Replacement candidates for abandoned or archived dependencies. Verify a candidate
     exists in the registry before naming it, and label these as judgment, not
     measurement.
   - For flagged install scripts: whether `npm ci --ignore-scripts` is viable for this
     project's build.

## Style for what you add

Write added prose the way a security report reads, and apply the same register to the
report addendum and the final reply alike — replies get pasted into tickets and reports
verbatim. State the finding, the datum behind it, and the action.

- Impersonal and declarative: no first or second person ("I ran the collector", "you
  should upgrade"), no contractions, no exclamation points.
- Active voice, with the subject matter as the actor: "upgrading to 1.19.0 clears all
  25 advisories", not "it is recommended that axios be upgraded".
- Objective: no intensifiers or subjective framing ("very", "significant",
  "fortunately"), and no guesses about why the project chose what it chose.
- Tense: past for what the audit did, present for the state of the dependencies,
  future for the consequences of acting or not.
- Constructive: a recommendation names the action and its cost, never a culprit.

If the `report-writing:writing-style` skill is available in the session, follow it —
it is the full version of this register.

The rendered report carries facts only. The interpretive rules below are instructions
to you, not content for the reader — do not copy them into the deliverable as caveats
or framing.

## Reading the report

- **Unassessable is not risk.** PyPI publishes no maintainer ACL and Go has no registry;
  those rows say what could not be known, not what is wrong.
- **The coverage table bounds every claim.** "No advisories" means "none among what was
  assessed" — check the assessed count before repeating a clean verdict.
- **Quote figures verbatim.** Do not re-derive, round, or embellish the report's
  numbers; every one is reproducible from the artifact.
- **Absence from the findings is not endorsement.** A dependency with no findings was
  measured against these criteria only.

## Rationalizations to reject

- "`gh` can give me maintainer counts faster than the collector." Measured wrong — repo
  contributors and registry publish rights are different populations.
- "No findings, so the dependencies are safe." Read the coverage table; on PyPI and Go,
  half the criteria are structurally unassessable.
- "The unassessable rows would just confuse the reader; I'll drop them." They are the
  boundary of every claim in the report. Dropping them turns partial coverage into a
  clean bill of health, which is the failure this skill was rebuilt to prevent.
- "The version is probably close enough." A range checked at latest-release and a
  lockfile-resolved version are different claims; the report labels which one it makes.
  Keep the label.

## When not to use

- License compliance auditing.
- Scanning the target's own source for vulnerabilities or secrets — this skill never
  reads dependency source, only registry, advisory, and repository metadata.
- Judging whether the project installs or builds. The audit is designed to work from
  nothing more than the dependency list — manifests and lockfiles — and never installs,
  builds, or executes anything. Broken installs and import-time breakage are out of
  scope, and worth saying so if the user seems to expect them.
- Ecosystems other than npm, PyPI, and Go; say the ecosystem is unsupported rather than
  improvising an audit for it.

讨论

基于 GitHub Discussions。登录 GitHub 即可参与讨论、点赞、订阅更新。