The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill uses dspy.PythonInterpreter({}).execute(expression) in multiple examples (found in SKILL.md and examples.md). This allows for the execution of arbitrary Python code from strings, which is a powerful but high-risk capability when exposed to an LLM.
[COMMAND_EXECUTION]: The SKILL.md file suggests running npx skills add lebsral/DSPy-Programming-not-prompting-LMs-skills --skill ai-do. This command fetches and executes code from a remote repository via the npx utility.
[EXTERNAL_DOWNLOADS]: The skill references several external dependencies and services:
Installation of the langchain-community package via pip.
Retrieval of data from a hardcoded IP address: http://20.102.90.50:2017/wiki17_abstracts.
External API calls to https://wttr.in/ via the requests library.
[PROMPT_INJECTION]: The skill facilitates an Indirect Prompt Injection surface (Category 8):
Ingestion points: The agents are designed to ingest data from untrusted external sources using tools like DuckDuckGoSearchRun, WikipediaQueryRun, and the search_wikipedia function.
Boundary markers: The provided code examples lack boundary markers or specific instructions for the LLM to ignore instructions embedded within retrieved search results.
Capability inventory: The agents are equipped with dspy.PythonInterpreter (arbitrary code execution), requests.get (network access), and potentially other tools.
Sanitization: No input validation or sanitization logic is presented for the content passed from the search tools to the interpreter or other agent modules.

ai-taking-actions