A single tampered weather sensor at Paris CDG airport netted a trader $20,000 on Polymarket, and experts warn AI-driven forecasting is heading toward a data-sabotage crisis

Someone pointed a hairdryer at a weather sensor at Paris Charles de Gaulle Airport, and the temperature jumped from 18 degrees Celsius to 22. That four-degree spike triggered a payout of $20,000 on Polymarket for whoever had bet on it 1.

The sensor was hit twice, on April 6 and April 15, 2026. On both days, the reading at CDG suddenly registered 22 degrees before dropping back to the actual spring temperature of around 18 2. French police and Météo-France are investigating. The suspected tool: a hand-held hairdryer or lighter applied directly to the instrument 3. On some of the affected days, more than $500,000 was staked on Paris temperature contracts on Polymarket, and three separate wallets collectively won more than $280,000 on a single day 3.

A $20,000 heist is not the story. The story is what almost did not catch it, and what will not catch the next one.

The only reason anyone noticed the CDG manipulation is that members of Infoclimat, a French climate nonprofit, were browsing the data and spotted the anomalies by chance 1. They flagged it to Météo-France, which filed a complaint with airport police 2. Behind that catch sits a structural safeguard: data assimilation. In traditional forecasting systems like the ECMWF Integrated Forecasting System, every incoming measurement is weighed against what the physical model says should be happening and against readings from nearby stations 1. If one station reports 22 degrees while its neighbors read 18, the system flags the outlier.

Now researchers at ECMWF are exploring whether forecasts can be produced directly from raw observations, skipping the assimilation step entirely 1. Other teams are combining weather station data with large language models and agentic AI to drive real-time, autonomous decisions during extreme events 1. These approaches, known in the field as data-driven models, consume raw observations more aggressively than any physics-based system, with less of the human judgment that caught the CDG fraud.

In an op-ed for MIT Technology Review, researchers from the Fraunhofer Heinrich Hertz Institute, ECMWF, and partner institutions lay out a risk ladder that climbs fast. At the bottom: the CDG case, one speculator rigging one station. One step up: coordinated traders biasing forecasts of renewable energy output to move wholesale electricity prices. At the top: a state actor manipulating one or many stations to trigger a false early warning or suppress a real one 1.

The technical gap between rungs is smaller than it sounds. Existing quality controls caught the CDG tampering because a single station reporting four degrees above its neighbors is statistically loud. But if someone remotely nudged the readings at many stations simultaneously, making each change small enough to look plausible on its own, current methods would struggle to distinguish coordinated manipulation from natural variation 1. And timing works against the defenders: careful data checks take hours or days, but forecasts go out on schedule 1.

The financial incentive to exploit this gap is already here. Polymarket, launched in 2020, lets users bet on real-world events including daily city temperatures 2. Institutional investors including Goldman Sachs use Polymarket data to inform their own trades 3. The CDG case proves the concept is not theoretical: physical access to one instrument produced a guaranteed return on a prediction contract. Polymarket has since stopped using the CDG sensor and switched to one at Paris-Le Bourget airport 3.

Every forecasting system has always been only as good as the observations feeding it. What is changing is how many attack surfaces those observations create. As weather prediction shifts from physics-based models that cross-check inputs against physical laws to neural networks that absorb raw sensor data wholesale, every weather station worldwide becomes a potential entry point for anyone with a motive and a heat source. The researchers' proposed safeguards, from station-level security to AI adversarial robustness tools to continuous human oversight along the entire data chain, all depend on one thing the field is actively trying to remove: humans in the loop 1.

The CDG fraud was caught because a human happened to be looking. The field is building toward a future where no human is positioned to look. The $20,000 was the cheapest proof of concept anyone will ever need.