The Agent Skills Directory

[PROMPT_INJECTION]: The skill is designed to process untrusted data from Discord messages, which represents an indirect prompt injection surface. To mitigate this, the instructions explicitly command the agent to treat all fetched content as data only and never interpret it as instructions or tool calls. The static analysis detections for instruction overrides are false positives, as they occur within security logic designed to detect and flag injections in the archived content rather than attempting to override the agent's behavior.
[COMMAND_EXECUTION]: The skill utilizes curl for downloading attachments but enforces strict safety wrappers. It requires that all filenames pass through sanitize_filename (which removes path traversal components and unsafe characters) and all URLs pass validate_url (which blocks private/internal IPs and limits downloads to specific Discord CDN domains).
[DATA_EXFILTRATION]: While the skill performs network requests via curl, these are limited to downloading content from trusted Discord infrastructure. The skill includes a redact_cdn_url function to ensure that ephemeral authentication tokens found in Discord URLs are not persisted in local logs or manifests, preventing potential credential exposure.
[DYNAMIC_EXECUTION]: The skill uses browser automation to extract data from Discord DMs. The JavaScript executed during this process is statically defined within the skill's reference files and is used solely for DOM extraction, posing no risk of executing arbitrary remote code.

discord-harvest