The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The template.md file provides a blueprint for new skills that includes the command sudo apt-get install. Encouraging the agent to use sudo facilitates acquiring administrative privileges that may not be necessary for the task.
[DYNAMIC_EXECUTION]: The core functionality of the skill involves the agent writing arbitrary script files (such as driver.mjs, driver.py, or smoke.sh) and then immediately executing them to interact with an application. This pattern of runtime code generation and execution creates a large attack surface.
[COMMAND_EXECUTION]: The instructions guide the agent to perform extensive system-level operations, including installing software via apt-get, managing background processes with tmux, and using xvfb-run to execute GUI applications in a headless environment.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze external, untrusted project data (such as README.md, package.json, and CI configurations) to determine how to build and run an application, creating a vulnerability surface for indirect injection.
Ingestion points: The agent surveys project files like README.md, package.json, Dockerfile, and Makefile in the 'Discover' phase.
Boundary markers: The instructions do not define boundary markers or provide warnings to the agent to disregard instructions found within these ingested files.
Capability inventory: The skill grants the agent capabilities to write files to the filesystem and execute arbitrary shell commands.
Sanitization: There is no mention of sanitizing or validating the contents of the project files before they are used to influence the generated driver code.
[EXTERNAL_DOWNLOADS]: The skill instructs the agent to download and install various packages from external registries (APT, NPM, PyPI) to satisfy project dependencies, which introduces external code into the execution environment.

run-skill-generator