The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The script scripts/finetune_model.py utilizes the trust_remote_code=True parameter within AutoTokenizer.from_pretrained and AutoModelForCausalLM.from_pretrained. This setting enables the execution of arbitrary Python code bundled with model files downloaded from remote repositories like the Hugging Face Model Hub, which can lead to code execution if the model repository is untrusted or compromised.
[PROMPT_INJECTION]: The skill implements a Retrieval-Augmented Generation (RAG) architecture in REFERENCE.md and scripts/setup_rag_pipeline.py, creating a surface for indirect prompt injection.
Ingestion points: Data is ingested from the local file system (e.g., ./knowledge_base/) into vector databases (ChromaDB and Pinecone) in REFERENCE.md and scripts/setup_rag_pipeline.py.
Boundary markers: The prompt templates in REFERENCE.md use basic instructional delimiters (e.g., "Answer the question using ONLY the provided sources") which are insufficient to prevent the agent from obeying malicious instructions embedded within retrieved document chunks.
Capability inventory: The skill includes powerful tools for model fine-tuning (scripts/finetune_model.py), serving (scripts/serve_model.py), and infrastructure management, increasing the potential impact of a successful injection.
Sanitization: There is no evidence of content sanitization or validation of the retrieved text before it is interpolated into the generation prompt.
[PROMPT_INJECTION]: The REFERENCE.md file contains a security filtering function check_safety which lists common prompt injection patterns like "ignore previous instructions" and "disregard all prior". While used defensively, these strings trigger security detectors for malicious instructional overrides.
[COMMAND_EXECUTION]: Multiple scripts within the skill (e.g., scripts/serve_model.py, scripts/benchmark_models.py, scripts/evaluate_model.py) are designed to be executed via the shell and perform operations such as managing Docker containers, spawning subprocesses, and modifying local system configurations.

llm-architect