The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process user-supplied input through LLM agents, which is an inherent surface for indirect prompt injection attacks.
Ingestion points: The run() and runner.run() functions in SKILL.md and various templates (e.g., templates/text-agents/agent-basic.ts) ingest untrusted text directly from user queries.
Boundary markers: While the basic templates lack explicit boundary markers, the skill provides a dedicated guide and template for implementing InputGuardrail and OutputGuardrail to detect and block malicious instructions.
Capability inventory: The skill demonstrates capabilities including network requests (via fetch in Cloudflare Workers templates), local command execution (via helper scripts), and tool calling (e.g., getWeatherTool, sendEmailTool).
Sanitization: The skill includes comprehensive examples for input validation, PII filtering, and content safety through its guardrail system (e.g., templates/text-agents/agent-guardrails-input.ts).
[COMMAND_EXECUTION]: The skill includes a shell script scripts/check-versions.sh used for development purposes to verify that dependencies are up to date. It uses standard npm view commands to check the registry for package versions, which is a legitimate and safe development utility.
[SAFE]: The skill demonstrates best-practice secret management by instructing users to store the OPENAI_API_KEY as an environment secret rather than hardcoding it. Additionally, it provides a secure architecture for voice agents that generates short-lived ephemeral session tokens on the server to prevent exposing primary API keys to the client-side browser.

openai-agents