qlora

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (high risk: 1.00). The skill's required training example explicitly calls datasets.load_dataset("tatsu-lab/alpaca", ...), pulling a public, user-contributed Hugging Face dataset that the agent ingests and uses to fine-tune the model, so untrusted third-party content can materially influence its behavior.