The Agent Skills Directory

[DATA_EXFILTRATION]: The skill instructions explicitly describe how to read local files using the file:/// URI scheme in the mediaUrl parameter for actions such as sendMessage, emojiUpload, and stickerUpload. This capability allows an agent to access sensitive local files (like SSH keys, environment variables, or credentials) and exfiltrate them by sending them as attachments or uploading them as media to a Discord server controlled by an attacker.- [PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection. It ingests untrusted data from external sources through the readMessages and searchMessages actions. Because the agent also possesses high-privilege capabilities such as timeout, ban, roleAdd, and deleteMessage, an attacker could place malicious instructions in Discord messages that the agent might inadvertently execute when processing those messages.
Ingestion points: Discord message content retrieved via readMessages and searchMessages (SKILL.md).
Boundary markers: None identified. There are no instructions for the agent to ignore or delimit embedded instructions in the messages it reads.
Capability inventory: The skill provides access to moderation tools (timeout, roleAdd), message management (deleteMessage, editMessage, pinMessage), and data transmission (sendMessage, stickerUpload, emojiUpload).
Sanitization: No evidence of sanitization or validation of the message content before processing.

discord