The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process content from a project's codebase, including package.json, Makefile, and documentation files. It specifically instructs the agent to 'Run project commands mentioned in docs to verify they work', which is an exploitable surface if these files contain malicious instructions. ● Ingestion points: Project structure, package.json, Makefile, README.md, CLAUDE.md. ● Boundary markers: Absent; there are no instructions to the agent to treat embedded commands with caution. ● Capability inventory: Bash(npm*), Bash(yarn*), Write(), WebFetch(). ● Sanitization: Absent; there are no checks to validate if the commands to be executed are safe or restricted beyond the tool definition.
[COMMAND_EXECUTION]: The skill uses allowed tools Bash(npm*) and Bash(yarn*) to execute code derived from the project codebase. While limited to package manager commands, these can be used to trigger arbitrary script execution if the project's configuration files are compromised.
[EXTERNAL_DOWNLOADS]: The skill utilizes WebFetch(*) to validate external links and references. This involves making network requests to arbitrary external URLs found within the documentation.

sync-docs