Output harvesting is the large-scale collection of an AI model’s responses so they can be reused as training data. In machine learning, this can be a legitimate form of model distillation: a smaller system learns by imitating the outputs of a larger one. In abuse scenarios, the same process becomes a way to extract capability without access to the model’s weights or internal code.
It matters in cyber security because it turns public APIs into an attack surface. An operator can send repeated, narrow, high-volume prompts, collect the answers, and train a rival model on the traces. Defenders look for signs such as repetitive requests, unusual account coordination, low prompt diversity, and concentration on a specific skill set. Common controls include rate limiting, abuse monitoring, account graph analysis, and prompt logging. Output harvesting does not copy a model perfectly, but it can still reduce the cost of building a competing system and expose valuable behavior patterns.