ai-improving-accuracy

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

• Third-party content exposure detected (high risk: 0.90). The SKILL.md and reference.md explicitly show loading public datasets from HuggingFace (e.g., load_dataset("squad", ...)) and using those dev/train sets for evaluation and optimizers, so untrusted third-party content is ingested and can directly influence optimization decisions and program behavior.