The Agent Skills Directory

[COMMAND_EXECUTION]: The installation script scripts/install.sh uses sudo apt-get install to install system-level dependencies like Graphviz on Linux, which involves executing commands with elevated privileges.
[EXTERNAL_DOWNLOADS]: The skill configuration and instructions direct the agent to download and install third-party code from the Agents365-ai GitHub organization via npx skills add and git clone operations.
[PROMPT_INJECTION]: The skill design presents an indirect prompt injection surface as it ingests and processes potentially untrusted external data (natural language descriptions and local source code) without documented sanitization. * Ingestion points: Natural language input and local codebase files (Python, JS/TS, Go, Rust) processed by scripts like pyimports.py and jsimports.py. * Boundary markers: Not present; the instructions do not specify any delimiters or instructions to ignore embedded commands within the processed data. * Capability inventory: The skill utilizes the Bash, Read, Write, Edit, and WebFetch tools, providing a wide range of actions that could be exploited by injected instructions. * Sanitization: There is no evidence of input validation or content escaping before the data is interpolated into the workflow.

drawio