The Agent Skills Directory

[METADATA_POISONING]: The skill contains deceptive and misleading metadata. It claims support for 'gpt-5-mini' and 'gpt-5-realtime', which are not currently released or supported models. Additionally, the skill includes a 'Last Verified' date of 2026-01-21, which is in the future. There is also a version discrepancy between the main documentation (v0.4.1/Zod 4) and the provided configuration files (v0.2.1/Zod 3).
[INDIRECT_PROMPT_INJECTION]: The skill provides templates for building agents that ingest untrusted data from external sources, such as API requests in the Next.js and Cloudflare Worker templates. This creates a surface for indirect prompt injection attacks.
Ingestion points: Untrusted data enters the agent context via request.json() in various API and Worker templates (e.g., templates/cloudflare-workers/worker-text-agent.ts).
Boundary markers: The templates do not consistently implement boundary markers by default, although a specific template (templates/text-agents/agent-guardrails-input.ts) is provided to demonstrate their implementation.
Capability inventory: The agents described in the templates have the capability to call tools, perform network operations via those tools, and delegate tasks to other agents.
Sanitization: The skill provides a dedicated template for input/output validation guardrails as a mitigation strategy, but these are not enabled by default in the simpler templates.

openai-agents