The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides an interface for agents to receive and process emails via webhooks or API polling as seen in scripts/check_inbox.py and references/WEBHOOKS.md. This establishes an attack surface where external actors can send malicious instructions to the agent.\n
Ingestion points: Incoming email body (text, html) and subject fields processed by the agent.\n
Boundary markers: No specific boundary markers or delimiters are implemented in the provided Python scripts to isolate untrusted email content; however, SKILL.md includes a section on 'Untrusted markers' as a recommended defense layer.\n
Capability inventory: The agent has the ability to read local files for attachments in scripts/send_email.py, make network requests to the AgentMail API, and interact with external services as shown in the GitHub integration example in references/EXAMPLES.md.\n
Sanitization: The skill provides documentation and a TypeScript template for a sender allowlist filter to drop webhooks from untrusted sources, but this is an optional user-implemented configuration.\n- [COMMAND_EXECUTION]: The provided scripts (scripts/send_email.py, scripts/check_inbox.py) use standard Python libraries to interact with the filesystem and network. While these are intended functionalities, they could be leveraged by an agent if it is successfully manipulated via an incoming email.\n- [EXTERNAL_DOWNLOADS]: The skill requires the installation of the agentmail and python-dotenv packages from standard registries. These are documented dependencies for the service.\n- [PROMPT_INJECTION]: A static analysis warning was triggered due to the presence of strings like 'Ignore previous instructions'. However, these appear exclusively within a security warning section in SKILL.md as examples of malicious inputs to guard against, and do not constitute an attempt to subvert the agent's instructions.

agentmail