The Agent Skills Directory

[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface due to the combination of untrusted data ingestion and system-level tool access. 1. Ingestion points: Data is loaded from external files via pd.read_csv('data.csv') and retrieved from databases using SQL queries. 2. Boundary markers: The skill instructions do not specify any delimiters or safety prompts to ignore instructions that might be embedded within the datasets. 3. Capability inventory: The skill utilizes the Read, Grep, Glob, and Bash tools, the latter of which allows for arbitrary shell command execution on the host environment. 4. Sanitization: There is no evidence of input validation, escaping, or sanitization of the data before it is processed or used in analysis steps.

data-analysis