The Agent Skills Directory

[PROMPT_INJECTION]: Indirect prompt injection surface identified. The skill allows agents to ingest and process data from external emails, which can contain adversarial instructions.
Ingestion points: references/webhooks.md and references/websockets.md (processing message.received event payloads).
Boundary markers: None identified in the instruction snippets; content is consumed as raw strings.
Capability inventory: SKILL.md includes capabilities to send and reply to emails, creating a potential loop for automated exploitation if the agent responds to malicious prompts.
Sanitization: No sanitization or validation of the email body is demonstrated in the examples.
[EXTERNAL_DOWNLOADS]: Instructs users to install the agentmail SDK from NPM and PyPI, along with support libraries like express and flask for webhook handling.
[COMMAND_EXECUTION]: Provides shell commands for dependency installation and suggests using the ngrok utility to expose local services.

agentmail