The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill's installation workflow defined in references/installation-workflow.md explicitly automates the execution of arbitrary shell scripts (bash setup.sh) downloaded from unvetted third-party GitHub repositories found during search. This capability allows for the execution of unverified and potentially malicious code within the user's local environment.
[EXTERNAL_DOWNLOADS]: The core functionality of the skill relies on fetching content and cloning repositories from unvetted third-party sources on GitHub using gh api and gh repo clone based on user-supplied search queries. This architectural design encourages the downloading of unvetted executable content from the open internet.
[COMMAND_EXECUTION]: The skill facilitates and automates high-risk system-level commands, including the use of package managers (npm install, pip install -r requirements.txt) and shell script execution from discovered repositories. This bypasses standard manual vetting procedures for third-party software installation.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it is instructed to 'Actually READ' and semantically analyze untrusted SKILL.md content from external repositories to perform scoring and evaluation.
Ingestion points: Untrusted content from third-party GitHub repositories is fetched via API and processed locally (e.g., stored in temp_skill.md) for semantic matching and quality assessment.
Boundary markers: The instructions lack explicit boundary markers or safety directives to isolate the external content, increasing the risk that malicious instructions embedded within a discovered skill could override agent behavior during the evaluation phase.
Capability inventory: The skill environment possesses extensive capabilities, including shell access (gh, bash), network operations, and file system modification permissions.
Sanitization: There is no evidence of sanitization, escaping, or validation of the external markdown files before they are processed by the agent's logic.

skill-finder