The Agent Skills Directory

[PROMPT_INJECTION]: The skill includes meta-instructions (e.g., 'Prefer retrieval over pre-training') designed to ensure the AI agent uses current documentation rather than outdated knowledge. These are benign and intended for grounding.
[DATA_EXPOSURE_AND_EXFILTRATION]: The documentation consistently advocates for safe secret management using wrangler secret, environment variables, and Pulumi secrets. It explicitly warns against anti-patterns like hardcoding credentials or caching sensitive environment objects in global scopes.
[UNVERIFIABLE_DEPENDENCIES_AND_REMOTE_CODE_EXECUTION]: The skill references official SDKs and well-known libraries (e.g., AWS SDK, Pulumi, Drizzle ORM, Puppeteer). These fall under trusted sources according to standard security rules. Remote code patterns documented (like curl API calls or sandbox.exec) are foundational components of the services being described.
[INDIRECT_PROMPT_INJECTION]: The reference covers services designed to ingest external data (Email Workers, AI Search indexing, Browser Rendering). While this creates an inherent attack surface, the documentation mitigates this by providing guidance on input validation, sanitization, and the use of prepared statements to prevent SQL injection.
[PRIVILEGE_ESCALATION]: No commands were found that attempt to acquire unauthorized system privileges. Usage of sudo or chmod in documentation is limited to standard troubleshooting and setup contexts for CLI tools.

cloudflare