The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill explains how to use experimental features like Codemode and Browser Tools, which enable an agent to generate and execute JavaScript code at runtime within an isolated sandbox. This allows the AI to perform complex logic and browser-based tasks dynamically.
[INDIRECT_PROMPT_INJECTION]: The skill provides patterns for ingesting data from untrusted sources, such as emails and webhooks. This creates a surface for indirect prompt injection attacks where malicious content within the external data could attempt to hijack the agent's logic.
Ingestion points: Handlers such as onEmail in references/email.md and onRequest in references/webhooks-push.md where external data enters the agent context.
Boundary markers: The documentation suggests verifying webhook signatures but does not explicitly show the use of prompt boundaries to wrap untrusted text.
Capability inventory: Agents built with this SDK have access to network operations (fetch), SQL databases, persistent state, and background task scheduling.
Sanitization: The skill mentions signature verification as a method to authenticate incoming data.

agents-sdk