The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill provides explicit instructions and command templates for the agent to download and execute an external package named agentic-awesome-skills from the public npm registry. This behavior allows for the execution of unverified binaries on the host system, which could perform arbitrary actions outside the skill's own source code.- [COMMAND_EXECUTION]: The orchestration logic depends on the agent executing shell commands for package installation and running the installer tool in both 'dry-run' and active modes.- [INDIRECT_PROMPT_INJECTION]: The skill is configured to ingest instructions from project-level files such as docs/users/workflows.md and data/workflows.json. If these files contain malicious payloads, the agent could be manipulated into performing unintended actions during the workflow execution.
Ingestion points: docs/users/workflows.md and data/workflows.json within the local environment.
Boundary markers: None specified; instructions are treated as a 'Source of Truth'.
Capability inventory: Includes shell command execution (npm exec), file system access, and invocation of other skills.
Sanitization: No methods for sanitizing or validating the content of these external workflow files are provided.- [METADATA_POISONING]: The skill's frontmatter includes a risk: none field. This is a deceptive claim that minimizes the significant risks associated with the dynamic execution of external code and the processing of untrusted input.- [EXTERNAL_DOWNLOADS]: The skill triggers network requests to the npm registry (registry.npmjs.org) to fetch and run the installer utility during the setup phase.

antigravity-workflows