The Agent Skills Directory

[DATA_EXFILTRATION]: The skill instructs the agent to search through sensitive environment files to detect project configuration. \n
Evidence: grep -r "openai\\|anthropic\\|google" .env* 2>/dev/null || true \n
Impact: Accessing .env files exposes potential API keys and other secrets to the agent's context, which is a significant data exposure risk. \n- [COMMAND_EXECUTION]: The skill utilizes several shell commands for project analysis and diagnostic testing. \n
Evidence: Uses grep for version checks, find and xargs for identifying SDK usage in source code, and curl to verify local streaming endpoints. \n- [EXTERNAL_DOWNLOADS]: The skill recommends using ecosystem-standard tools for codebase updates and migrations. \n
Evidence: npx @ai-sdk/codemod upgrade \n
Note: This utilizes the official Vercel codemod package for automated upgrades, which is a well-known service. \n- [METADATA_POISONING]: The skill includes misleading information in its metadata and implementation examples. \n
Evidence: The displayName "AI SDK by Vercel (v5)" suggests official origin, and the documentation references fictional AI models such as "gpt-5", "claude-opus-4.1", and "llama4". \n- [INDIRECT_PROMPT_INJECTION]: The skill analyzes external codebase files, creating a vulnerability surface where instructions embedded in project files could influence agent behavior. \n
Ingestion points: Analyzes package.json, source code files (.ts, .tsx), and configuration files (.env*). \n
Boundary markers: No delimiters or protective instructions are included to isolate content from potential embedded prompts. \n
Capability inventory: Shell command execution (grep, find, npm, npx) and local network requests (curl). \n
Sanitization: No sanitization is applied to the data ingested from the project environment.

ai-sdk-expert