The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The workflow in Step 4 instructs the agent to install tools using "pip, npm, apt, or direct download" from GitHub repositories identified through search. This facilitates the execution of arbitrary, unvetted code from external sources at runtime.
[COMMAND_EXECUTION]: The skill requires the agent to execute shell commands to install and run third-party utilities with user-supplied files. This capability can be exploited if the chosen tool is malicious or if the agent is directed to use unsafe command-line arguments.
[EXTERNAL_DOWNLOADS]: The skill promotes downloading and running software from the internet without a fixed whitelist of trusted vendors, relying on easily-manipulated heuristics like "stars" and "last commit" dates to assess safety.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection. It requires the agent to ingest and analyze untrusted data (GitHub READMEs and documentation) which can contain malicious instructions.
Ingestion points: GitHub repository content, README files, and project documentation.
Boundary markers: None; the agent is not instructed to use delimiters or ignore embedded instructions when parsing repo content.
Capability inventory: Full shell access for package installation (apt, pip, npm) and execution of arbitrary binaries.
Sanitization: None; the skill lacks any validation or escaping mechanisms for the content it processes.

github-gem-seeker