openai-agent-sdk-skill

Pass

Audited by Gen Agent Trust Hub on Apr 7, 2026

Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONREMOTE_CODE_EXECUTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The skill provides instructions to install software packages from the NPM registry, including @openai/agents, zod, and @modelcontextprotocol/server-filesystem.- [COMMAND_EXECUTION]: Includes shell commands for project initialization, environment configuration, and running MCP servers via npx.- [REMOTE_CODE_EXECUTION]: Describes the use of hostedMcpTool which connects to an external URL (https://gitmcp.io/openai/codex) to provide tool capabilities.- [DATA_EXFILTRATION]: Contains examples for retrieving agent context and handling API keys through environment variables.- [PROMPT_INJECTION]: Includes standard instructions (e.g., 'You are a helpful assistant'). Note that instructions recommend using 'gpt-5-nano', which does not correspond to a publicly known model.- [INDIRECT_PROMPT_INJECTION]: The skill documents agents that process untrusted user input, creating a potential attack surface. \n
  • Ingestion points: User messages and conversation history in SKILL.md. \n
  • Boundary markers: Absent in code examples. \n
  • Capability inventory: Sensitive tools such as delete_file and codeInterpreterTool are defined in SKILL.md. \n
  • Sanitization: Documents the implementation of InputGuardrail and OutputGuardrail components to validate and filter content.
Audit Metadata
Risk Level
SAFE
Analyzed
Apr 7, 2026, 08:53 AM
Security Audit — agent-trust-hub — openai-agent-sdk-skill