The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The skill instructions in SKILL.md explicitly guide the agent to prompt the user for escalated network access and to rerun commands with escalated permissions when necessary for deployment tasks.
[DYNAMIC_EXECUTION]: The documentation in the references/sandbox/ directory provides extensive APIs for creating and managing edge containers that execute arbitrary shell commands and Python code (e.g., sandbox.exec() and ctx.runCode()).
[REMOTE_CODE_EXECUTION]: A pattern in references/sandbox/patterns.md demonstrates downloading and executing a script from an external source using curl -fsSL https://code-server.dev/install.sh | sh. This is a well-known method for installing development tools in a remote environment.
[INDIRECT_PROMPT_INJECTION]: The skill describes processing untrusted external data in references/email-workers/, where the agent is instructed to parse incoming MIME emails and attachments. The documentation correctly notes risks like address spoofing but the attack surface remains present by design.
[EXTERNAL_DOWNLOADS]: Numerous reference files list dependencies on official packages from well-known services and organizations, including Cloudflare, OpenAI, Vercel, and Pulumi. These are used for legitimate functionality such as AI SDK integration and infrastructure-as-code management.

openai-cloudflare-deploy