The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill directs users to install the genai-tk toolkit from a third-party GitHub repository (https://github.com/tclatos/genai-tk@main) rather than a verified package registry.
[REMOTE_CODE_EXECUTION]: The toolkit includes support for SmolAgents and a python_repl tool that allow for the generation and execution of Python code. Although Docker isolation is recommended, the toolkit permits 'local' execution in the host process.
[COMMAND_EXECUTION]: The CLI system (genai-tk.main.cli) uses dynamic imports to load command classes defined in YAML configuration files, which allows for arbitrary code loading if the configuration is modified.
[DATA_EXFILTRATION]: The system manages sensitive API keys for LLM providers in .env files. Agents with filesystem_tools access can read these files, and when paired with network-capable tools like web_search or browser_use, this creates a risk of data exfiltration.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it processes untrusted data from web searches and local documents. The mandatory evidence chain includes:
Ingestion points: External data enters through RAG ingestion (rag add-files) and agent tools like web_search and browser_use.
Boundary markers: There are no explicit delimiters or instructions provided to distinguish between system instructions and retrieved data content.
Capability inventory: The toolkit enables subprocess execution, Python REPL usage, filesystem access, and network requests.
Sanitization: No content sanitization or instruction filtering is implemented for data retrieved from external sources.

genai-tk