The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill provides patterns for generating and executing Python code at runtime using the exec() function.
Evidence: In SKILL.md, the create_analysis_function demonstrates wrapping LLM-generated strings into a function using exec(wrapped_code, {"pd": pd, "np": __import__("numpy")}, local_vars). This allows the agent to execute arbitrary code generated by the LLM.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted data from various sources which could contain malicious instructions designed to influence the code generation process.
Ingestion points: The skill ingests data from uploaded CSV/Excel files and user-provided natural language queries (found in SKILL.md under the FastAPI and Streamlit integration sections).
Boundary markers: The provided code examples do not show explicit boundary markers or system instructions to ignore instructions embedded within the processed dataframes.
Capability inventory: The skill has the capability to execute code via exec() and shell commands via the Bash and Write tools listed in the SKILL.md frontmatter.
Sanitization: The PrivacyAwareAnalyzer class implements basic keyword-based filtering (e.g., blocking "show name"), but does not provide comprehensive sanitization against adversarial prompt injection within data values.
[EXTERNAL_DOWNLOADS]: The skill instructions include the installation of several external Python packages from public registries.
Evidence: Commands such as pip install pandasai openai matplotlib seaborn plotly openpyxl xlrd sqlalchemy psycopg2-binary pymysql are recommended for full functionality.

pandasai