Malware detection without training data — AI now classifies threats by reasoning about behavior

Researchers built a system that identifies types of malware by asking multiple large language models to analyze code behavior, then combining their answers — instead of training on labeled examples. This means security teams can classify new or unknown malware threats without waiting for labeled datasets or manual feature engineering.

Most malware detection requires humans to first label thousands of examples, then train a system on them — a slow bottleneck when threats evolve faster than labeling can keep up. This approach skips that step by using AI models that can reason about code behavior directly, which means security analysts could start classifying novel threats immediately. The catch is nobody knows yet whether these AI decisions are reliable enough in practice or whether they're just very confident guesses.