The Agent Skills Directory

[DYNAMIC_EXECUTION]: The ConstructionRiskAssessor class in SKILL.md includes a load_models(path) method that utilizes joblib.load(). This function is built on Python's pickle module, which is inherently insecure and can execute arbitrary code if it is used to load a maliciously crafted model file from a user-provided or untrusted path.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process external construction data from CSV, Excel, and JSON formats to perform predictions.
Ingestion points: Data is loaded via pd.read_csv("project_history.csv") and the assistant is instructed to "Accept data in common formats: CSV, Excel, JSON" in instructions.md.
Boundary markers: There are no explicit delimiters or warnings to the agent to ignore instructions that might be embedded within the data fields of the construction logs or project histories.
Capability inventory: The skill possesses the ability to read and write to the filesystem (pd.ExcelWriter, joblib.dump, pd.read_csv) and output information to the user.
Sanitization: While the prepare_features method performs standard ML preprocessing such as encoding and handling missing values, it does not include sanitization or filtering to prevent the agent from interpreting text data as instructions.

risk-assessment-ml