The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill recommends installing vendor-owned packages agentmail and agentmail-toolkit from standard registries like PyPI and NPM.
[COMMAND_EXECUTION]: The documentation provides standard setup commands including pip install and npm install for managing dependencies.
[DATA_EXFILTRATION]: The skill manages communication data via the AgentMail API. A potential security risk is identified in references/examples.md (Pattern 7), which demonstrates saving email attachments using the unsanitized filename attribute from the message payload, which could lead to path traversal vulnerabilities if implemented without validation.
[PROMPT_INJECTION]: The skill demonstrates a surface for indirect prompt injection where untrusted external data influences agent behavior. 1. Ingestion points: Raw email bodies are ingested via webhooks in SKILL.md and references/examples.md. 2. Boundary markers: The example code lacks delimiters or instructions to ignore embedded commands within the email text. 3. Capability inventory: The agent is equipped with tools to send, reply to, and list messages, providing an exploitation path if the agent is compromised. 4. Sanitization: There is no evidence of sanitization or filtering of the email["text"] content before it is interpolated into the prompt for the LLM.

agentmail