The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill explicitly whitelists and promotes the use of the pickle library in SKILL.md for saving and loading analysis results. The pickle module is known to be vulnerable to arbitrary code execution during deserialization if the input data originates from an untrusted source.
[INDIRECT_PROMPT_INJECTION]: The skill's primary function is to ingest and analyze external data sources (CSV, JSON, pickle, or experiment logs) provided as the $0 argument. This represents a vulnerability surface where maliciously crafted data could attempt to influence the agent's logic or code generation during the multi-round review process.
Ingestion points: Data is loaded via the load_data function in scripts/stat_summary.py and referenced as the primary input in SKILL.md.
Boundary markers: The skill lacks explicit instructions or delimiters to isolate the data content from the agent's instructional context.
Capability inventory: The skill has the ability to write to the file system (via the --output flag in scripts) and use the pickle library for serialization.
Sanitization: There is no evidence of sanitization or filtering of the input data to prevent embedded instructions from being processed as part of the research goal analysis.

data-analysis