The Agent Skills Directory

[COMMAND_EXECUTION]: The skill configuration in SKILL.md requests the Bash tool with cmd:* permissions. This allows for the execution of arbitrary shell commands. While this permission is used to run the bundled ML scripts, it represents a high-privilege capability that could be misused if the agent is misled.
[PROMPT_INJECTION]: The skill processes external datasets, creating a surface for indirect prompt injection attacks.
Ingestion points: Untrusted data is ingested by scripts/train_model.py and scripts/preprocess_data.py as part of the automated training workflow.
Boundary markers: The prompt instructions do not include delimiters or specific guidance to ignore instructions embedded within the datasets.
Capability inventory: The agent has access to Bash(cmd:*), Read, and Write tools, which provide powerful execution capabilities if a dataset contains malicious instructions.
Sanitization: The provided boilerplate scripts lack logic to sanitize or validate dataset contents against potential embedded instructions.

training-machine-learning-models