T03 · Remote Payload Retrieval and Execution
- Location
SKILL.md:24- Finding
Mutable Remote Installer Is Downloaded and Executed Directly by a Shell
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
The skill is coherent and not deceptive, but it asks agents to run mutable remote installers and npm scaffolding commands with user-level filesystem access.
Review carefully before installing. Prefer pinning exact versions of `@oia-ai/oia-skill` and `@oia-ai/oia-fresh`, avoid pipe-to-shell installation, and initialize only in a new or version-controlled directory so generated files can be reviewed or reverted.
SKILL.md:24Mutable Remote Installer Is Downloaded and Executed Directly by a Shell
SKILL.md:31Unpinned npm Package Is Downloaded and Executed Without Confirmation
声明描述的是一个“项目初始化”技能,核心行为应是搭建 Deno + Fresh 应用并执行相关验证流程。但实际代码仅是一个本地安装器,用于把技能文档复制到 Claude 的技能目录中。其主要目的与声明明显不同:实际代码围绕技能安装,而非项目脚手架初始化。虽然安装脚本可能是该技能包的一部分,但就所给代码片段而言,它执行了未声明的本地文件系统安装行为,同时完全没有实现声明中的主要能力,因此构成明显不匹配。
curl -fsSL https://deno.land/install.sh | sh downloads and immediately executes a remote script in the user's shell, which is a high-risk pattern. If the remote source, transport, DNS, or hosting is compromised, this becomes instant arbitrary code execution with the user's privileges.
- `deno --version`:要求 Deno 2.x。未安装时引导安装:
- Windows:`winget install Denoland.Deno`
- macOS / Linux:`curl -fsSL https://deno.land/install.sh | sh`
- `npx --version`:确认 Node / npx 可用(仅用于执行 init 脚本,项目运行不依赖 Node)。
### 2. 生成脚手架
The README instructs users to execute npx @oia-ai/oia-skill without pinning a specific version, which means the code fetched and executed can change over time. If the package is later compromised, typo-squatted, or a malicious version is published, users may execute attacker-controlled code directly during installation.
This global install command also uses npx without an exact version, creating a supply-chain risk where whatever version is current at execution time will run on the user's machine. Because it installs into a user-global Claude skills directory, a compromised package could persist and affect multiple projects rather than a single workspace.
The skill workflow instructs execution of npx @oia-ai/oia-fresh init without version pinning, which is especially risky because it scaffolds and runs project initialization code fetched at invocation time. In this skill context, that behavior is more dangerous than a passive dependency reference because the whole purpose of the skill is to have an agent execute the command automatically, increasing the chance of blind remote code execution from a compromised package release.
The skill metadata/manifest references the same package workflow without version pinning, reinforcing a pattern of executing mutable registry content. Manifest-level guidance can influence automated tooling and users to run unsafe bootstrap commands repeatedly.
The skill instructs use of an unpinned npm package via npx, which executes remote code at install/runtime from the latest published version. If the package is compromised, typo-updated, or the maintainer account is hijacked, users may run attacker-controlled code during project initialization.
This command again references @oia-ai/oia-fresh without an exact version, causing execution of whatever version is current in the registry at the time. That creates supply-chain risk because initialization tools typically have filesystem write access and can run arbitrary postinstall or scaffold logic.
Installing this skill package via npx -y @oia-ai/oia-skill without a version pin allows arbitrary code from the latest npm release to run. Even if intended for convenience, it exposes users to package takeover or malicious update risks.
The global install variant compounds the same unpinned execution risk and broadens persistence because it installs into a user-wide location. A compromised package could affect multiple projects or future invocations on the system.
Running npx -y @oia-ai/oia-fresh@latest init <项目名> explicitly tracks the latest release, which is a classic supply-chain hazard. Because scaffold generators write files and may execute setup logic, compromise of the upstream package can directly lead to arbitrary code execution and poisoned project contents.
The retry command still uses @latest, so adding proxy environment variables does not reduce the underlying remote-code execution risk from an unpinned package. In fact, it may make users more likely to execute the command automatically once connectivity is restored.
The markdown instructs the agent to confirm directory handling, but it also states that using the current folder will 'write/overwrite template files'. Because this can modify existing user files, the skill description should include a clearer user-facing warning about overwrite risk before proceeding.
npx commands without a version suffix (e.g. @1.0.0) create a rug-pull risk if the upstream server is compromised and publishes a malicious update.
npx commands without a version suffix (e.g. @1.0.0) create a rug-pull risk if the upstream server is compromised and publishes a malicious update.
npx commands without a version suffix (e.g. @1.0.0) create a rug-pull risk if the upstream server is compromised and publishes a malicious update.
The markdown directs the agent to start a background server and repeatedly poll localhost for verification, but it does not present this as a user-visible warning or consent point. Starting background processes and making network requests, even to localhost, can affect the user's system state and should be disclosed.
The script's docstring presents usage information only in Chinese, and the user-facing success message and trigger guidance are also Chinese-only. This can violate a language/locale policy when a skill forces a specific language without user opt-in or documented regional scope.
The natural-language description is written entirely in Chinese, which can impose a language/locale expectation on users without any opt-in or alternative language indicated in this file. Under the policy, language constraints should either offer user choice or be clearly justified as region-specific.