The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill instructs the agent to execute a command to install external code: npx skills add https://gitcode.com/Lingxi-HandsOn/gallery.git --skill publish-work-to-gallery. This facilitates the download and execution of instructions from a repository not managed by a verified trusted organization.
[EXTERNAL_DOWNLOADS]: The skill triggers the download of external resources from gitcode.com, a platform that is not recognized as a trusted organization or well-known service in the predefined security scope.
[INDIRECT_PROMPT_INJECTION]: The skill defines a vulnerability surface for indirect prompt injection:
Ingestion points: User-supplied content, including sandbox URLs (work_link), ECS addresses (api_link), and work descriptions (summary), are accepted as input and processed by the submit.mjs script (SKILL.md).
Boundary markers: The skill lacks explicit delimiters or instructions to the agent to treat this user-provided data as untrusted, increasing the risk that the agent may interpret embedded instructions as its own (SKILL.md).
Capability inventory: The agent environment has high-privilege capabilities including shell command execution (npx skills add), network operations via fetch, and read/write access to the local file system (SKILL.md, submit.mjs).
Sanitization: Although submit.mjs implements basic domain regex and string length validation, these checks do not sanitize the content against malicious instructional payloads designed to manipulate the agent's behavior during the submission process.

huawei-cloud-activity-submit