The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill documentation and instructions include a command that fetches a remote shell script and executes it directly: 'curl -fsSL https://raw.githubusercontent.com/PadaliyaSavan88/api-forge/main/install.sh | bash'. This pattern allows for the execution of arbitrary unverified code on the host system.\n- [EXTERNAL_DOWNLOADS]: During installation, the skill downloads executable content from an external GitHub repository ('PadaliyaSavan88/api-forge') that is not a recognized or well-known technology provider.\n- [COMMAND_EXECUTION]: The skill uses shell commands like 'npx tsc --noEmit' and 'npx skills add' to validate or install project components. These commands execute code that may be influenced by local or remote project data.\n- [PROMPT_INJECTION]: The skill analyzes 'package.json' and the 'src/' directory to determine the project state (SKILL.md). Mandatory Evidence Chain: 1. Ingestion points: reads 'package.json' and local source files. 2. Boundary markers: Absent; no delimiters or ignore-instructions are used. 3. Capability inventory: can write files (scaffolding templates) and execute 'npx' commands. 4. Sanitization: Absent; content is processed directly to influence agent logic. This creates a surface for indirect prompt injection where malicious project data could override agent instructions.

backend