The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill implements a model loading mechanism using pickle.load() within the MLModelRetrainer.load_model method in SKILL.md. This function deserializes data from a file path provided as an argument. The Python pickle module is inherently unsafe because it can execute arbitrary code embedded within the pickled data during deserialization. An attacker who can influence the file path or supply a malicious .pkl file could achieve arbitrary code execution in the environment.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process external construction data from various formats such as Excel, CSV, and JSON to monitor for data drift and trigger retraining. This creates a surface for indirect prompt injection where malicious instructions hidden in the dataset could influence the agent's behavior.
Ingestion points: The skill accepts project data, file paths, and parameters as specified in instructions.md and uses pandas to read these files.
Boundary markers: There are no explicit delimiters or instructions within the prompts to ignore or isolate embedded instructions in the processed data.
Capability inventory: The skill possesses filesystem access permissions (claw.json) and uses them to read/write model and data files.
Sanitization: While the code uses basic data cleaning like .fillna(0), it lacks security-focused sanitization to detect or filter prompt injection payloads in text-based data fields.

ml-model-retrainer