The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses the runcomfy CLI tool to execute image editing tasks. The analysis confirms that inputs are passed as a JSON string (--input), which prevents shell injection vulnerabilities by ensuring the model prompt is not evaluated by the host shell.
[EXTERNAL_DOWNLOADS]: The skill relies on the @runcomfy/cli Node.js package and interacts with model-api.runcomfy.net. These are official resources belonging to the vendor (agentspace-so / RunComfy), and the skill explicitly whitelists these domains for output downloads.
[CREDENTIALS_UNSAFE]: The skill documents the use of RUNCOMFY_TOKEN for authentication and notes that local tokens are stored in ~/.config/runcomfy/token.json with appropriate 0600 permissions. It correctly advises users to use environment variables in CI/CD environments, which is a standard safe practice for secret management.
[INDIRECT_PROMPT_INJECTION]: As an image editing skill, it processes external image URLs. The documentation explicitly acknowledges the risk of image-based prompt injection and advises treating external URLs as untrusted data, demonstrating high security awareness.
[SAFE]: No malicious patterns, obfuscation, or unauthorized data exfiltration behaviors were detected. The skill's behavior aligns perfectly with its stated purpose of providing a CLI wrapper for a specific AI model.

flux-kontext