The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The script scripts/dpo-training.py utilizes the trust_remote_code=True parameter within the AutoModelForCausalLM.from_pretrained and AutoTokenizer.from_pretrained methods. This setting allows the execution of custom Python code defined within a remote model's repository, posing a risk if the model source is untrusted or compromised.
[EXTERNAL_DOWNLOADS]: The setup guide in rules/local-ollama-setup.md provides an installation command that pipes a remote script from https://ollama.ai/install.sh directly into the shell. While Ollama is a well-known service, this pattern is inherently risky as it executes remote code without prior local inspection.
[DATA_EXFILTRATION]: The configuration script scripts/create-lora-config.md includes commands to search for model-related variables within .env files using grep. Because environment files frequently contain sensitive credentials and secrets, accessing them programmatically increases the risk of unintended data exposure.

llm-integration