Large language models (LLMs) are increasingly used to predict future events by reading the news. But news can be fake. A person who knows only the forecasting question can publish a plausible fake article to steer the prediction.
Researchers tested this attack on real stock and prediction-market questions. The attacker had no access to the model or its search system, just published one article before the forecast. They measured how often the forecast flipped.
Once the fake article was retrieved, it reversed 74% to 82% of forecasts that originally went the other way. Even the strongest defenses were bypassed, and accuracy did not reveal the harm. Forecasting systems that search the open web are especially exposed.