The Agent Skills Directory

SAFE (INFO): The skill is a set of educational templates that prioritize security best practices. For example, agent-human-approval.ts demonstrates mandatory human-in-the-loop (HITL) authorization for sensitive operations like delete_account and send_email.- SAFE (INFO): Includes dedicated safety templates in agent-guardrails-input.ts and agent-guardrails-output.ts to detect and mitigate prompt injection, jailbreak attempts, and PII leaks.- SAFE (INFO): Secure architectural patterns for voice agents are demonstrated in api-realtime-route.ts, showing how to generate ephemeral session tokens server-side to avoid exposing the OPENAI_API_KEY to the client browser.- COMMAND_EXECUTION (LOW): The maintenance script scripts/check-versions.sh executes the npm view command to verify package versions against the registry. This is a standard utility and poses no security risk.- INDIRECT_PROMPT_INJECTION (LOW): While the templates ingest untrusted user input, the package includes extensive mitigation strategies (Guardrails, HITL, and structured output validation) to secure the agent's reasoning and capability surfaces.

openai-agents