The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill documents how to process email content, which is an untrusted external data source that could contain malicious instructions. The skill mitigates this by providing patterns for using clear boundary markers (delimiters), instructional framing, and allow lists to restrict senders. It also includes an authorization matrix to define which actions require human confirmation.
[PROMPT_INJECTION]: The documentation contains examples of prompt injection attacks (e.g., 'Ignore your previous instructions') within its security threat model. These are provided for educational purposes to demonstrate risks and are not instructions for the agent itself to follow.
[CREDENTIALS_UNSAFE]: The skill provides defensive code patterns that include regular expressions for detecting potential secrets (such as AgentMail and OpenAI keys) in outbound emails to prevent accidental data leakage. It does not contain any hardcoded credentials.
[COMMAND_EXECUTION]: Code snippets demonstrate the use of the agentmail library and the svix package for webhook verification. These are standard integrations for the intended functionality of managing email inboxes and processing events.
[DATA_EXFILTRATION]: The skill discusses data isolation patterns, such as using dedicated inboxes and multi-tenant pods, to prevent cross-agent data exposure and ensure clear audit trails.

agent-email-patterns