The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: Automated security scanners have flagged the skill's primary file (SKILL.md) as malware (FileRepMalware). Additionally, multiple URLs linked in the skill, including the author's GitHub Pages domain jeffallan.github.io, are flagged as malicious and appear on security blacklists. These links are found in the frontmatter and documentation sections.
[COMMAND_EXECUTION]: The deployment instructions in references/deployment-optimization.md include a function that executes shell commands via subprocess.run to perform GGUF model conversion. The script path for this conversion is derived from the LLAMA_CPP_PATH environment variable, which could be exploited if an attacker controls the environment.
[INDIRECT_PROMPT_INJECTION]: The skill is heavily focused on ingesting and processing external data for training and fine-tuning models, which is a significant vector for indirect prompt injection.
Ingestion points: Processes dataset files in .jsonl, .json, and .parquet formats using the datasets library across SKILL.md and references/dataset-preparation.md.
Boundary markers: The instructions lack specific guidance or code for isolating training data to prevent the fine-tuned model from adopting malicious instructions embedded in the datasets.
Capability inventory: The skill uses network-enabled clients (OpenAI, Hugging Face), writes to the local file system (checkpoints, adapters), and executes subprocesses.
Sanitization: While the skill includes dataset validation for schema and quality (e.g., filtering for 'As an AI' boilerplate), it does not sanitize or filter for adversarial instructions that could poison the model.

fine-tuning-expert