The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs users to install several packages from the ElevenLabs ecosystem, including the CLI (@elevenlabs/cli), the JavaScript SDK (@elevenlabs/elevenlabs-js), and browser/React clients (@elevenlabs/client, @elevenlabs/react). It also references a widget script hosted on a well-known CDN (https://unpkg.com/@elevenlabs/convai-widget-embed). These resources originate from a well-known service and are considered safe.
[INDIRECT_PROMPT_INJECTION]: The skill enables agents to process untrusted user voice and text transcripts, creating a surface for indirect prompt injection where a user might attempt to override agent instructions. Ingestion points: User transcripts captured via the conversational AI session (e.g., onUserTranscript events). Boundary markers: The documentation explicitly references a built-in prompt_injection guardrail in the platform_settings.guardrails configuration to detect instruction overrides. Capability inventory: Agents have access to external tools including server-side webhooks, browser-side JavaScript execution, and telephony operations (outbound calls/transfers). Sanitization: The sanitize property in dynamic_variables allows for redacting sensitive values from transcripts and tool responses.
[DYNAMIC_EXECUTION]: The platform supports "Client Tools" which are explicitly designed to execute JavaScript in the user's browser. While this is a core functionality for UI interaction and navigation, it represents a dynamic code execution vector for content generated by the agent's LLM.

agents