An AI now spots a brain bleed before a doctor can even see it

In partnership with

Here is a number that should stop you cold: when a stroke cuts off blood to your brain, you lose about 1.9 million neurons every minute it goes untreated. Not “eventually.” Right now, this minute. Doctors have a grim little phrase for it — time is brain — and the whole game of emergency stroke care is a race against a clock that is shredding you while everyone runs.

This week, at a health-tech event in Israel, doctors said something that would have sounded like science fiction a few years ago: AI is now flagging life-threatening brain bleeds in seconds — before a physician can even spot them on the scan.

That is today’s story. It is one of the most genuinely hopeful uses of AI going. And — because this is itscybernews and there is always a catch — it comes with a trapdoor that the same researchers, in the same country, discovered years ago and that nobody has fully closed.

Marvel first. As always.

🧠 The wonderful part: a second set of eyes that never blinks

Picture the old way. You arrive at the ER slurring your words. You wait for a CT scan. The scan is taken, then it joins a queue of images waiting for a radiologist — who may be asleep at 3 a.m., or reading for four hospitals at once, or simply thirtieth in line. Every step is a few more minutes. Every few minutes is a few million more neurons.

Now picture the new way. The instant the CT scanner finishes, an AI model reads the image before it even reaches the queue. If it sees the signature of a bleed or a large clot, it fires an alert straight to the on-call stroke team’s phones — the neurologist, the specialist who removes clots, everyone at once — with the images attached.

This isn’t a pilot or a promise. Two systems — Viz.ai and Aidoc — already do exactly this in hospitals around the world, cleared by the FDA. And the numbers are the good kind of unbelievable:

• Viz.ai alerts the stroke team in under 30 seconds from the scan, and reads the image itself in about 2 minutes — against 15 to 60 minutes for a traditional workflow.

• A large real-world study across many hospitals found AI cut the time from a patient’s arrival to the specialist being contacted by nearly 40 minutes.

• One hospital saw the variation in alert times collapse — from a wild swing of over two hours down to a tight seven-minute window. Consistency, when your brain is on the clock, is its own kind of miracle.

Here’s what 40 minutes actually buys, in a real case doctors described: a man suddenly lost control of his right side, was rushed in, scanned within minutes, and the AI flagged the stroke instantly. Treatment started right away. The next day he was sitting up and talking. In another, the AI caught a dangerous clot, the patient was transferred immediately, and — in the phrase every neurologist dreams of — woke up on the table completely neurologically intact.

And the accuracy is real: these systems hit 90–95% for spotting large clots and bleeds, matching or slightly beating human radiologists on speed-critical calls. Nobody’s replacing the doctor. The AI is the tireless junior who never looks away from the screen, so the senior human can act sooner.

It doesn’t stop at strokes. This same week brought AI reading ECGs while hiding the patient’s identity, AI giving ALS patients their voice back, AI matching cancer patients to personalized care. The pattern is everywhere: the machine handles the first, frantic, time-critical look — and hands a calmer, faster decision to the human.

So what could possibly go wrong with a system this good? The answer is unsettling, and it starts with a question almost nobody asks: if a doctor is trusting the picture — can you trust the picture?

🩻 The catch: the scan you’re betting your life on can be forged

Years ago, researchers at Ben-Gurion University in Israel — the same country now celebrating AI stroke detection — asked exactly that question. Their answer became one of the most disturbing security demonstrations in medicine.

They built a piece of malware called CT-GAN. Using the same family of AI that powers deepfakes, it could reach into a CT or MRI scan as it traveled through a hospital’s network and add a realistic cancerous tumor that was never there — or erase a real one that was. In close to real time. Seamlessly.

Then they tested it on actual radiologists. The results are the stuff of nightmares:

• When doctors didn’t know the scans were tampered with, they diagnosed 99% of the fake injected tumors as real cancer, and 94% of scans with cancer secretly removed as “healthy.”

• Even after being told the images might be forged and put on high alert, they were still fooled 60% of the time by added tumors and 87% by removed ones.

• The cancer-detection AI that doctors use as a backstop? It was fooled by the fakes 100% of the time.

Sit with that. A forged medical image can send a healthy person into chemotherapy — or tell a dying one they’re fine. And the trust we’re building in AI-read scans makes the forged picture more dangerous, not less, because the whole point of these systems is that we act on them faster, with fewer humans double-checking.

It gets more real than a lab. You don’t even need exotic malware to weaponize the scan pipeline — you just need to turn it off. And attackers do, constantly:

• In 2024, a ransomware attack on Ascension, one of America’s largest health systems, knocked out electronic records and forced ambulances to divert to other hospitals — delaying tests and treatment across dozens of facilities.

• In January 2025, Frederick Health in Maryland was forced offline by ransomware, diverting ambulances and delaying care.

• In October 2025, Heywood Healthcare in Massachusetts pulled its network down and diverted ambulances — with radiology and lab services explicitly limited.

Now stack that against time is brain. When ransomware takes a hospital’s CT scanner or its network offline, the stroke patient doesn’t get the two-minute AI alert. They get put in an ambulance and driven somewhere else — burning the exact minutes that AI was invented to save. Studies of hospitals near ransomware attacks have found emergency cases spiking and survival rates dropping. The cyberattack two towns over can cost a stranger their brain.

The miracle and the vulnerability are the same system. A faster, more automated, more networked path from scanner to specialist is wonderful right up until someone tampers with the path — or cuts it.

Your prompts are leaving out 80% of what you're thinking.

When you type a prompt, you summarize. When you speak one, you explain. Wispr Flow captures your full reasoning — constraints, edge cases, examples, tone — and turns it into clean, structured text you paste into ChatGPT, Claude, or any AI tool. The difference shows up immediately. More context in, fewer follow-ups out.

89% of messages sent with zero edits. Used by teams at OpenAI, Vercel, and Clay. Try Wispr Flow free — works on Mac, Windows, and iPhone.

Start flowing free

🛡️ The good news: the fixes are known, and mostly boring

The reassuring truth about CT-GAN is that it exploited something dull and fixable: hospital imaging networks that moved scans around unencrypted and unsigned, on the trusting assumption that anything inside the hospital was safe. That assumption is the bug. The medicine here is the same kind of unglamorous hygiene that stops most attacks.

• Sign and encrypt the images. If every scan is digitally signed the moment the scanner creates it, a tampered image fails the signature check before any human or AI ever reads it. The standards to do this exist; hospitals just have to turn them on.

• Don’t trust the internal network by default. CT-GAN worked because it assumed the hospital’s own network was a safe zone. Modern “zero trust” security drops that assumption — every device and message has to prove it belongs.

• Keep the human in the loop for the big calls. AI as the fast first look, a human confirming anything life-altering, is exactly the right shape. The 100%-fooled detector is a warning: never let only the machine decide.

• Treat the scanner like critical infrastructure — because it is. The ransomware defenses that protect a hospital’s billing system protect its CT scanner too: offline backups, network segmentation so one infected PC can’t reach the imaging system, and a tested plan for running when the network is down.

None of this is exotic. It’s seatbelts, not rocket science.

✅ What to actually do

If you or someone you love may face a stroke (which is all of us):

1. Learn the signs and call emergency services instantly — BE-FAST: sudden Balance loss, Eyesight changes, Face drooping, Arm weakness, Speech trouble → Time to call. The AI only helps if you get to the scanner. Your speed still matters most.

2. It’s fair to ask if your hospital uses AI stroke tools. Systems like Viz.ai and Aidoc are a genuine advantage in an emergency. Knowing which nearby hospitals are “comprehensive stroke centers” is worth doing before you ever need one.

If you work in or run a healthcare organization:

1. Ask your imaging team one question: are our scans signed and encrypted in transit? If the answer is “no” or “not sure,” you have a CT-GAN-shaped hole. This is a fixable configuration, not a rebuild.

2. Segment the imaging network and back it up offline. The goal is that a ransomware infection on a front-desk PC can never reach — or freeze — the CT scanner.

3. Rehearse downtime. The hospitals that weather an attack are the ones that already practiced running without the network. For stroke care, have a pre-agreed diversion and manual-read plan so minutes aren’t lost to confusion.

The takeaway

We are living through something quietly extraordinary: a machine that can look at a picture of your bleeding brain and raise the alarm faster than the best-trained human in the building — buying back the minutes that are, very literally, you. That’s not hype. It’s saving lives in hospitals right now, and it’s worth being amazed by.

But the same leap that makes the scan faster and more trusted also makes the scan worth attacking. A picture we act on instantly is a picture worth forging — and a pipeline we depend on completely is a pipeline worth cutting. The Israeli researchers who showed us the miracle this week are from the same country whose researchers showed us the forgery years ago. Both are true at once. That’s the whole story of technology, really.

The good news is that the locks are known and dull: sign the images, don’t trust the network blindly, keep a human on the life-and-death calls, and treat the scanner like the critical machine it is. Time is brain. Let’s make sure the only thing racing that clock is the medicine — not the malware.

Stay curious, stay safe.

— itscybernews