The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill implements several patterns for loading and saving models using libraries that are vulnerable to arbitrary code execution through unsafe deserialization. Specifically, the 'FeaturePipeline.load' method in 'references/feature-engineering.md' uses 'pickle.load()', and multiple components in 'references/pipeline-orchestration.md' use 'joblib.load()', both of which can execute arbitrary code when processing a malicious file. Additionally, 'references/training-pipelines.md' utilizes 'torch.load()' for checkpoints, which is also based on pickle and carries similar risks.- [INDIRECT_PROMPT_INJECTION]: The skill designs pipelines that ingest untrusted data from external sources (e.g., Parquet files, S3/GCS buckets) for feature engineering and training, creating an indirect prompt injection surface. While the skill includes data quality checks via Great Expectations, it lacks security-focused sanitization to prevent adversarial data from influencing model behavior or exploiting vulnerabilities in data-processing libraries.- [COMMAND_EXECUTION]: The skill provides a shell script 'launch_distributed.sh' that executes 'torchrun' with configurable arguments. While intended for distributed training, this pattern allows for shell command execution within the agent's environment.

ml-pipeline