The Agent Skills Directory

[REMOTE_CODE_EXECUTION] (HIGH): The skill facilitates the installation and execution of 'custom nodes' via comfy node install <name>. Since custom nodes are arbitrary Python scripts, an attacker could trigger the installation of malicious code by providing a compromised node name or registry entry.
[EXTERNAL_DOWNLOADS] (HIGH): The command comfy model download --url <url> allows the agent to download files from any web address. This can be exploited to pull malicious binaries or scripts onto the local system.
[CREDENTIALS_UNSAFE] (MEDIUM): The tool manages and stores sensitive API tokens for CivitAI and Hugging Face (civitai_api_token, hf_api_token). While necessary for gated models, these are stored in plaintext configuration files (config.ini), making them targets for exfiltration.
[COMMAND_EXECUTION] (MEDIUM): The skill provides high-level commands for system-level operations, including launching background servers (comfy launch --background), stopping processes, and modifying the environment, which could be abused to disrupt the host system.
[INDIRECT_PROMPT_INJECTION] (LOW):
Ingestion points: Processes workflow.json and snapshot.json files.
Boundary markers: None detected in documentation or command definitions.
Capability inventory: comfy node install-deps --workflow can trigger automated package installations based on file content.
Sanitization: No evidence of sanitization or validation of the workflow files before processing.

comfy-cli