The Agent Skills Directory

[COMMAND_EXECUTION]: The skill defines a calculate tool that uses the built-in Python eval() function to evaluate user-provided expressions. This is a critical security risk as it allows for arbitrary code execution if an attacker or a subverted agent passes malicious Python strings.
[REMOTE_CODE_EXECUTION]: The agent configuration includes the PythonREPLTool from the langchain_experimental package. This tool provides the AI agent with a persistent Python shell, enabling it to execute any code it generates. This capability can be easily abused to perform unauthorized file system operations, network scanning, or data exfiltration.
[PROMPT_INJECTION]: The skill implements Retrieval-Augmented Generation (RAG) using PyPDFLoader and WebBaseLoader to ingest external content.
Ingestion points: External data is loaded from document.pdf and web sources into the agent context.
Boundary markers: The prompt templates for the RAG chain and Agent do not employ delimiters or 'ignore' instructions to prevent the agent from obeying malicious instructions embedded within the retrieved text.
Capability inventory: The agent possesses powerful capabilities including arbitrary code execution (eval, PythonREPLTool) and network access (requests, DuckDuckGoSearchRun).
Sanitization: No sanitization or validation of the retrieved context is performed before it is interpolated into the LLM prompt.

langchain