The Agent Skills Directory

[Prompt Injection] (SAFE): The skill includes defensive templates (e.g., agent-guardrails-input.ts) specifically designed to detect and block prompt injection attempts. No malicious instructions or bypass markers were found in the templates or documentation.
[Data Exposure & Exfiltration] (SAFE): The skill provides explicit guidance and code templates for securing API keys. It demonstrates how to generate short-lived ephemeral tokens on the server (Next.js) to avoid exposing the primary OPENAI_API_KEY to the client side. No hardcoded credentials or unauthorized data access patterns were detected.
[Unverifiable Dependencies & Remote Code Execution] (SAFE): All dependencies in templates/shared/package.json are legitimate libraries (@openai/agents, zod). The version check script scripts/check-versions.sh uses standard npm view commands for maintenance and does not download or execute untrusted remote content.
[Indirect Prompt Injection] (LOW): While agents process untrusted user data in templates like agent-structured-output.ts and agent-streaming.ts, the skill consistently implements and advocates for the use of Zod for strict schema validation and instruction-based guardrails to sanitize inputs and outputs.
[Dynamic Execution] (SAFE): The templates use standard JavaScript/TypeScript patterns for agent orchestration. No use of eval(), exec(), or other dangerous dynamic code execution methods was found.

openai-agents