The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill utilizes an exec() call within scripts/data_transformer.py to run pandas-based code for data transformation. This code is generated by the AI agent based on the structure of user-provided data files.
Mitigation: The skill implements a multi-layered sandbox to restrict the execution environment. This includes:
A Keyword Blacklist that blocks dangerous functions and attributes (e.g., open, import, socket, subprocess, os.system).
An AST Whitelist that allows only specific, safe Python syntax nodes.
Restricted Builtins that remove access to dangerous global functions like __import__, eval, and exec from within the transformation context.
Resource Limits including an execution timeout (10 seconds) and recursion depth limits.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted external data files (CSV, Excel, JSON). This creates an attack surface where malicious content embedded in data files could attempt to influence the agent's behavior during the analysis phase or when generating descriptive annotations for the charts.
Evidence Chain:
Ingestion points: Data is read via scripts/data_parser.py from user-specified file paths.
Boundary markers: The skill instructions in SKILL.md define a clear contract for the transformation code, but do not specify specific delimiters for data values during LLM analysis.
Capability inventory: The skill has file read/write permissions and uses exec() to perform data operations.
Sanitization: The skill performs column name normalization and type conversion, and the DataTransformer provides safety filtering for the code generation step.

smart-charts