The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs the agent to clone the NousResearch/pokemon-agent repository from GitHub, which is a third-party source not included in the trusted vendor lists.
[REMOTE_CODE_EXECUTION]: The instructions command the agent to install the cloned package in editable mode (pip install -e .[pyboy]) and execute its server component, which leads to the execution of unverified remote code.
[COMMAND_EXECUTION]: The skill uses ssh to create a reverse tunnel to the public service localhost.run. This is used to expose the local game server port (9876) to the public internet, potentially bypassing local firewall security and exposing the host to external access.
[DATA_EXFILTRATION]: While intended for a live dashboard, the use of a public SSH tunnel creates a pathway for data exposure or unauthorized remote access to the local machine's services.
[PROMPT_INJECTION]: The skill presents an attack surface for indirect prompt injection (Category 8) as it processes unverified external data.
Ingestion points: RAM state and screenshots retrieved from the game emulator (SKILL.md).
Boundary markers: None are used to separate game data from instructions.
Capability inventory: Includes remote package installation, shell command execution via ssh, and background process management (SKILL.md).
Sanitization: No validation or filtering is performed on game state data or visual inputs before they are passed to the agent's logic.

pokemon-player