The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from Discord (messages, filenames, embeds) which could contain malicious instructions. This risk is well-mitigated through a mandatory 'staging' step where the agent inventories assets before downloading, a flag_suspicious() function that detects common injection patterns, and a clear security policy that prevents the agent from interpreting the archived content as instructions.\n- [COMMAND_EXECUTION]: The skill provides Bash templates for downloading and organizing files. Security is maintained through strict input validation: validate_url() restricts downloads to a hardcoded Discord CDN allowlist and blocks internal/private IP ranges to prevent SSRF, while sanitize_filename() prevents path traversal and shell-unsafe characters from being used in local file writes.\n- [DATA_EXFILTRATION]: While the skill handles ephemeral Discord CDN tokens during the download process, it includes a redact_cdn_url() function to ensure these tokens are never logged or persisted in manifests. It does not send user data or Discord tokens to unauthorized third-party servers.\n- [PROMPT_INJECTION]: A static analysis flag was raised for an instruction override pattern in references/folder-structure.md. Investigation confirms this is a false positive; the text is part of a documentation example demonstrating how the skill's manifest schema logs detected injection attempts found in Discord metadata.

discord-harvest