AI-Generated Medical Deepfakes Fool Radiologists, Raising Network Security Risks

The healthcare sector is facing a novel and sophisticated threat vector: the medical deepfake. A recent study published in the journal Radiology, led by Dr. Mickael Tordjman of the Icahn School of Medicine at Mount Sinai, has demonstrated that synthetic X-ray images generated by AI models like ChatGPT and RoentGen are now virtually indistinguishable from genuine patient data. This development transcends mere academic curiosity, signaling a profound vulnerability in the digital infrastructure of global healthcare systems.

The empirical data from the study is sobering. When seventeen radiologists from twelve hospitals across six countries were presented with a mix of 264 real and synthetic images, their spontaneous detection rate was a mere 41%. Even when explicitly warned that the dataset contained fakes, their accuracy only improved to 75%, leaving a significant margin for error. This suggests that the human element—traditionally the final safeguard in clinical diagnosis—is no longer sufficient to verify the authenticity of medical imagery in an era of generative AI.

“

Detection accuracy among these models fluctuated between 57% and 85%.

Perhaps more alarming is the failure of AI to police itself. The study tested four leading large language models (LLMs): OpenAI’s GPT-4o and GPT-5, Google’s Gemini 2.5 Pro, and Meta’s Llama 4 Maverick. Detection accuracy among these models fluctuated between 57% and 85%. Notably, GPT-4o, the very model used to create some of the deepfakes, was unable to identify all of its own synthetic outputs. This "AI blindness" complicates the development of automated verification tools, as the generative capabilities of these models appear to be outstripping their discriminative counterparts.

From a cybersecurity perspective, the implications are twofold. First, there is the risk of "data poisoning" or network injection. If a threat actor gains unauthorized access to a hospital’s Picture Archiving and Communication System (PACS), they could theoretically replace or augment real patient records with synthetic images. Such an attack could be used to manipulate clinical outcomes, extort healthcare providers, or cause widespread operational paralysis. Unlike traditional ransomware, which encrypts data, this form of attack undermines the integrity of the data itself, making it impossible for clinicians to trust the information on their screens.