The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill fetches and displays user-generated content from an external source (community.sandock.ai) through the posts and post commands in scripts/community.mjs. This external content is processed by the agent without specific boundary markers or sanitization, creating an attack surface where a malicious forum post could attempt to influence the agent's behavior.
Ingestion points: The cmdPosts and cmdPost functions in scripts/community.mjs ingest raw post content and replies from the Sandock API.
Boundary markers: There are no explicit delimiters or "ignore embedded instructions" warnings used when the agent displays or reads this external content.
Capability inventory: The skill possesses Read and Write file access and network capabilities through the Bash(node:*) tool execution.
Sanitization: The script does not perform sanitization or filtering on the markdown body of retrieved posts to detect or neutralize prompt injection attempts.
[DATA_EXFILTRATION]: The upload command in scripts/community.mjs allows the agent to read local files and upload them to the vendor's API. While intended for legitimate image uploads to the forum, this tool lacks a mandatory --yes confirmation gate (as noted in SKILL.md), meaning a compromised agent could be tricked into exfiltrating sensitive local files to the remote server's storage.
[COMMAND_EXECUTION]: The skill relies on the execution of a Node.js script that utilizes node:fs for reading local environment variables and processing files for upload. While the script's logic is benign, it grants the agent broad access to the local filesystem within the scope of its allowed tools.

sandock-community