The Agent Skills Directory

EXTERNAL_DOWNLOADS (MEDIUM): The setup instructions require users to install the agentmail-mcp package via npx -y or pip install. These packages are hosted on public registries and are not from the listed trusted organizations.
DATA_EXFILTRATION (LOW): The skill requires an AGENTMAIL_API_KEY and processes sensitive email data. While expected for its purpose, it involves transmitting this data to the agentmail.to domain, which is not in the whitelist of trusted exfiltration domains.
COMMAND_EXECUTION (LOW): The skill requires executing shell commands for setup and operation, including npx, pip, and the agentmail-mcp binary.
Indirect Prompt Injection (LOW): The skill reads external data (emails) which could contain malicious instructions designed to manipulate the AI agent.
Ingestion points: Tools such as get_thread, list_threads, and get_attachment ingest content from external email messages.
Boundary markers: No boundary markers or specific safety instructions are provided to the agent for handling untrusted email content.
Capability inventory: The agent has capabilities like send_message, reply_to_message, and delete_inbox which could be misused if the agent obeys instructions within an email.
Sanitization: There is no evidence of content sanitization before the data is passed to the AI assistant.

agentmail-mcp