To screen for lung cancer, radiologists typically view hundreds of images from a single CT scan. With this new AI model, Google can generate an overall lung cancer malignancy prediction and identify subtle malignant tissue, or lung nodules, which are often difficult to see. The AI also factors in previous scans, which can help reveal the growth rate of suspicious tissue.

To test the model, Google asked its AI to examine 45,856 chest CT screens. It compared the results against six board-certified radiologists. In these first studies, Google’s AI detected five percent more cancer cases than the radiologists. It also reduced false-positive exams by more than 11 percent.

The model needs additional clinical research and testing before it can be deployed, but Google says the initial results are encouraging. The company also notes that only two to four percent of eligible patients in the US are screened for lung cancer. Like its other AI-based disease detection tools, Google hopes this one might make early detection more accessible.

To test out how effective the attack could be, the researchers conducted a blind study that asked radiologists to diagnose conditions based on CT lung scans—some of which were altered using the malware. When presented with scans that featured fake cancers nodules, the radiologists came back with a cancer diagnosis 99 percent of the time. When the malware was used to hide real cancer nodules, radiologists issued a clean bill of health 94 percent of the time.

Even when the radiologists were made aware that the scans were being altered, they still struggled to make a correct diagnosis. When they were given a second set of images with a warning that some had been changed, the medical professionals were still tricked into thinking computer-generated nodules were real 60 percent of the time. When the malware was used to remove nodules, 87 percent of the readings incorrectly determined the patient was healthy. The humans put through the test shouldn’t feel too bad, though—screening software used to confirm diagnoses fell for the malware’s tricks every single time.

The good news is the malware was created by security researchers and not malicious actors, so this particular tool isn’t likely to appear in the wild. But it should raise some red flags for medical professionals. Hospitals have been a target of cyber attacks before, but the stakes are usually more immediate: Ransomware locks up systems until a fee is paid. An attack like the one laid out by the researchers would be more insidious and could create distrust in essential systems.