The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill describes an architecture where agents ingest untrusted user input and possess the capability to execute powerful tools, creating a surface for indirect prompt injection.
Ingestion points: Untrusted data enters the agent context via the run method for text agents and session connections for voice agents, as shown in SKILL.md and templates/text-agents/agent-basic.ts.
Boundary markers: The templates provide natural language instructions to guide agent behavior (e.g., in templates/text-agents/agent-handoffs.ts), which serve as basic guidance but do not provide cryptographically secure isolation from untrusted input.
Capability inventory: The templates demonstrate tools with high-impact capabilities such as send_email, process_refund, and delete_account (specifically in templates/text-agents/agent-human-approval.ts).
Sanitization: The skill effectively promotes the use of Zod schemas for strict parameter validation and structured outputs, which significantly reduces the risk of malformed or malicious data reaching tool execution logic.

openai-agents