In a new study published in Science Translational Medicine on Wednesday, researchers say they have uncovered how Epstein-Barr virus launches immune responses that lead to the inflammation and nervous system damage seen in people with multiple sclerosis.
“It’s very nice now to be able to understand more about the underlying mechanisms of how EBV likely causes MS,” said lead author Kjetil Bjornevik, an assistant professor of epidemiology and nutrition at the Harvard T.H. Chan School of Public Health. The findings, researchers hope, could help with the development of EBV vaccines or antiviral medications that could prevent or manage MS symptoms without the major side effects of commonly used immunosuppressants.
Syed Rizvi, the director of the Multiple Sclerosis Center of Rhode Island, who was not involved in this study, said the new findings help advance MS research toward more precise approaches. “When you’re developing drugs, targeted drugs, every little step, every little molecule, every little antigen is a game changer,” he said.
Hello there from the final day of the Alzheimer’s Association International Conference. This is our last edition of this pop-up newsletter, but if you somehow haven’t tired of me, you can join me as well as my colleagues Damian Garde and Katherine MacPhail tomorrow to recap AAIC and discuss how the research presented here fits into the broader direction of the field. You canregister for the virtual event here. It’s at 10 a.m. Eastern, 3 p.m. here in the U.K.
With next year’s AAIC set for Chicago, and as this nation descends into full World Cup mania in the coming hours, I’ll end by saying thanks for following along with me here in London.
How blood tests could reshape the future of identifying dementia
Traditionally, an Alzheimer’s diagnosis comes after a brain scan or spinal tap, or perhaps some cognitive tests administered by a behavioral neurologist. The tests can be burdensome, and specialist capacity is limited.
But research presented throughout the conference indicated how the field is moving in new directions, finding ways to make testing much more accessible, and offering more nuanced results that go beyond saying whether someone has Alzheimer’s or not.
One study detailed here looked at whether these tests could be used in the primary care setting. Alzheimer’s experts say it’s crucial for more doctors to be able to diagnose the condition, particularly with the availability of new treatments that are more beneficial the earlier they can be used. Wait times for neurologists can extend for months, if not over a year.
WASHINGTON — Senate health leader Bill Cassidy (R-La.) grilled a Trump nominee for a key pandemic preparedness role over past comments in which he questioned vaccines, in a heated Senate confirmation hearing on Wednesday.
“Why would you repeat those damn lies? Because that destroys trust,” Cassidy said at one point to the nominee, Sean Kaufman, rapping his hand on the dais.
Cancer will touch 92% of people, either directly or through a close family member, the World Health Organization’s latest annual report on cancer warns. One in five people will develop the disease themselves. The report also found widespread global inequities in access to prevention, diagnosis, treatment and care.
This global view of cancer occurrence, diagnosis, and treatment, comes at a time when some in industry feel oncology is undergoing some dramatic technological transformations. Companies positioned to help spark change in how we meet the growing challenge of cancer around the world include those that provide direct to consumer genomic (e.g. Grail) or MRI (e.g. Prenuvo) testing for the disease. But whole new disciplines are likely to emerge, such as PreOncology, a new service that offers very early, comprehensive, cancer screening through primary care physicians (PCPs). PreOncology launched in Florida July 1.
“The number of oncology cases is not declining, unlike cardiology and other conditions,” Jose Barreau, MD, one of PreOncology’s co-founders, told Inside Precision Medicine. “Primary care doctors in particular, can use help with these patients.”
It is estimated there will be 20.6M cases of cancer and 10M deaths from the disease every year going forward. The number of cancer patients worldwide is expected to rise to nearly 35M cases by 2050. The report emphasizes inequities and points out that in developed countries, 85% of patients with breast or childhood cancers will survive at least five years. But in poorer countries that figure drops to less than 30%.
Approximately 38% of cancers can currently be prevented by avoiding risk factors, such as tobacco and alcohol, eating well and exercising, and implementing existing other “evidence-based prevention strategies,” WHO said.
But the rate of cancer and cancer deaths can also be reduced through early detection and appropriate treatment and care of patients: Many cancers have a very high cure rate if diagnosed early and treated appropriately.
The most common new cancer cases in 2024 were in: lung, breast, colon and rectum, prostate, skin (non-melanoma), and stomach. While the most common causes of cancer death last year were: Lung, colon and rectum, liver, breast, and stomach. All of which can be detected early using advanced diagnostics, but they are almost almost only screened for in older adults.
The incidence of cancer rises dramatically with age, but it is also a leading cause of death among children and adolescents. WHO reports that it is estimated that each year approximately 400,000 children and adolescents develop cancer. The rise of certain cancers in young adults, such as colorectal cancer and lung cancer in non-smoking, younger women, has been particularly alarming.
Cancer screening is still largely based on age and averages, and many cancers have no routine screening at all. But early screening proponents, such as PreOncology have integrated advanced technologies and proprietary AI to establish ways to accurately identify cancers as early as stage I, or the Alpha stage. This is when they are usually curable by the simplest means, and chemotherapy or other harsh treatments may be avoided.
PreOncology “looks at each person’s unique risk profile, from genetics and family history to lifestyle, to create a personalized monitoring plan,” the company said in a recent press release. An oncologist-led Signal Board reviews any concerning findings and helps guide physicians on the next steps.The goal is to detect cancer at the earliest point where something impactful can be done and when there are still the most options.
Cancer is the first or second leading cause of death, depending on which statistics you use.
The start-up hasa risk engine drawing on ~3.1M de-identified patient records from the studies behind national screening guidelines. Its lung model alone was modeled on 611,000+ people. The program’s whole genome sequencing component covers 61 cancer-related genes. Ambry genomics does the sequencing. The program also comprises whole-body MRIs and a variety of liquid biopsy type tests, CT-DNA, circulating tumor cells.
“There is a lot of emerging technology and early detection coming to market. Some of it is not FDA-approved, but it’s being advertised to physicians, and we can take advantage of that,” Barreau said.
To tie it all together, he added, “We have the deep medical oncology expertise to interpret and understand all this data in the context of any patient’s specific circumstances.”
Will such new approaches, along with new treatments such as cancer vaccines, be the basis for a new age of oncology? Let’s hope, as the rise in global prevalence of cancer continues.
A first-in-human Phase I trial has established preliminary safety and efficacy data for a reformulated version of thiostrepton, a mitochondrially targeted covalent PRX3 inhibitor, in patients with mesothelioma and other malignancies associated with malignant pleural effusion. The results were published in Nature Communications.
Thiostrepton was originally characterized decades ago and has since accumulated substantial preclinical evidence of anticancer activity, but its poor solubility and manufacturing complexity had previously precluded clinical development. The researchers undertook both the mechanistic characterization of the compound and the chemistry work required to bring it to a first-in-human setting.
“The goal was to take this drug called thiostrepton, which has been studied for 70 years, and formulate it in a way where it could be delivered to human beings for the first time ever, and use it as an anti-cancer drug,” explained principal investigator Brian Cuniff, PhD, associate professor at the University of Vermont.
Study population and design
Rather than restricting enrollment strictly to a single tumor histology, the trial enrolled patients on the basis of a shared clinical phenotype—malignant pleural effusion—which is most commonly, though not exclusively, associated with mesothelioma. This design allowed inclusion of a small subset of patients with metastatic lung and colorectal cancers alongside the mesothelioma-predominant cohort.
“Our primary goal is in the treatment of mesothelioma, but from a utilization and development standpoint, we’re very interested in testing it in other cancers,” Cuniff said.
Fifteen patients were enrolled, consistent with a dose-escalation, first-in-human design. “Around 30% of those patients we did see tumor reduction, and then we had 67% disease control,” Cuniff said, clarifying that disease control included patients whose “tumor didn’t grow for a period of time, or their tumor shrunk a little bit.”
A completed Phase II study is expected to provide a larger dataset on efficacy later this year.
Mechanistic rationale: PRX3 inhibition
The therapeutic rationale centers on thiostrepton’s activity as a covalent inhibitor of peroxiredoxin 3 (PRX3), a mitochondrial antioxidant protein. Preclinical work from Cuniff’s group established that irreversible PRX3 inhibition impairs mitochondrial bioenergetics and increases oxidative and metabolic stress selectively within tumor cells.
“It binds to the PRX3 protein irreversibly, and inactivates that protein, and that protein is really important for keeping the mitochondria of our cells… functioning correctly,” Cuniff said. “In tumor cells, we’re essentially taking away the ability for the tumor cell to make energy, and we’re also increasing waste products that are toxic to the tumor cells.”
A key element of the preclinical dataset is a reported therapeutic window between malignant and normal cell populations. “Our evidence shows that this same effect is not occurring in normal cells, and they can tolerate that inhibition much more than tumor cells,” Cuniff said.
Beyond direct cytotoxicity, investigators describe an additional immunomodulatory component to the drug’s activity, distinguishing it mechanistically from both chemotherapy and immune checkpoint-based approaches. “It’s not an immunotherapy, it’s not a chemotherapy,” Cuniff said. “It’s basically a cytotoxic immunomodulator. It kills cells, but it also has some activity against the immune system, which we think is important for its overall activity.”
Formulation and regulatory considerations
Historically, thiostrepton’s clinical translation was limited by poor aqueous solubility and manufacturing challenges. To address this, the development team formulated the compound into a micellar solution to enable solubilization and systemic delivery, and scaled manufacturing to meet regulatory-grade standards.
“We were able to manufacture the drug in a way that would allow it for the delivery to human beings, and then we developed a formulation that would allow for safe and effective delivery of the drug,” Cuniff said. The trial was conducted in the United Kingdom under MHRA oversight, the agency’s equivalent of the U.S. FDA.
Clinical context and unmet need
Cuniff situated the findings within a treatment landscape that has seen limited innovation. Standard-of-care chemotherapy for mesothelioma remained largely unchanged for roughly three decades until the 2021 approval of an immunotherapy-based frontline regimen—the most recent major regulatory advance for the disease.
“To only have two drugs essentially be approved in a 30-, 40-year period is unlike any other cancer,” Cuniff said, attributing the slow pace of development primarily to the disease’s rarity rather than to scientific tractability. “There’s not a ton of value in developing in mesothelioma because [of] the low patient population… although it really needs development. These patients, you know, they die very quickly. There’s not a lot of options, so we need new drugs.”
With Phase II data anticipated by year’s end, the investigators intend to further define thiostrepton’s efficacy profile in mesothelioma while continuing to evaluate its applicability across other PRX3-dependent tumor types.
Research shows fluorescence lifetime imaging microscopy (FLIM) combined with a form of artificial intelligence (AI) known as deep learning can predict if a person has lung cancer-related EGFR mutations with no need for genetic testing or tissue staining.
As reported in the journal Cancer Research, the AI model was able to achieve 96.6% accuracy on a standard diagnostic test and could also distinguish between the two most important subtypes of the EGFR mutation, which matters because they respond differently to treatment and carry different survival outlooks.
“With this research we are able to take a single fluorescent image and can very accurately predict whether the mutation is present. We can do this with a single image, without special stains and without gene sequencing,” co-lead author Ahsan Akram, MBChB, PhD, a professor at the Institute for Regeneration and Repair, University of Edinburgh, told Inside Precision Medicine.
“All the cells in our body, to some degree are capable of emitting light when excited by the correct type of laser light, what FLIM does it takes this emitted light and measures the time taken for the fluorescence to be emitted. The FLIM signal therefore captures a snapshot of the metabolic activity through an ‘optical fingerprint’.”
Lung cancer is the second most common cancer in the U.S. and by far the deadliest. For many lung cancer patients, particularly non-smokers, the key question is whether their tumor carries a mutation in a gene called EGFR, encoding a protein that drives cancer cell growth. If it does, they can receive effective targeted drugs, but looking for these mutations currently requires genetic testing. This is commonly PCR or next-generation sequencing, both of which are slow, expensive, and result in the tumor tissue sample being lost after testing.
The team used FLIM to scan lung tissue samples from 85 patients and trained a deep learning model (DenseNet-169) to classify each sample as EGFR-mutant or normal. The model achieved an area under the receiver operating characteristic curve score, a standard statistical measure of diagnostic accuracy, of 0.966, outperforming all previously published methods that rely on conventionally stained tissue images. The model also distinguished between the two most common EGFR mutation subtypes, an exon 19 deletion and an exon 21 point mutation.
“The machine learning was able to extract features that are specific to the EGFR mutation, and although we don’t exactly know what these are, when we test these on samples it has never seen before we saw remarkable accuracy in the prediction,” says Akram.
Although the microscopes needed to carry out FLIM are not cheap, this technique has several advantages that counteract this including speed of testing and also allowing the sample material to be reused for other tests as destructive staining is not needed.
“Current pathways require specialized labs and next generation sequencing. These are time consuming and costly. It can take weeks to have an answer. Here, as we are taking an image using laser light the process can take minutes,” he says.
“As we are shining laser light on the sample, without any stains, this is a completely nondestructive process. The tissue sample is intact following the image and then can be used for whatever else is needed. This is particularly important in lung cancer diagnostics as often we run out of tissue to do the full suite of tests we require in a cancer diagnostic workup.”
The authors acknowledge the current FLIM imaging and analysis process is somewhat time-consuming, taking around 1-2 hours per sample, but note that faster acquisition methods are in development. If validated in larger, more diverse patient cohorts, this type of testing could shorten the amount of time from biopsy to a patient receiving targeted treatment.
“We are at the stage of compelling proof of concept with strong performance on tissue samples. The next essential step is prospective clinical validation, testing the approach on samples collected in real time within clinical pathways, and demonstrating that it performs consistently and integrates practically into NHS laboratory workflows,” says Akram.
“In parallel, we are actively exploring the extension of this platform to other cancer types and additional targetable mutations, and investigating how FLIM can be integrated into existing clinical imaging infrastructure.”
For most new parents, that’s an interaction—you get the baby’s attention, and the baby babbles back—that is expected to happen. For Sierra, that never happened with her son, Travis. Born six weeks early, Travis spent about a week in the NICU. To get out of the NICU, he had to pass a newborn hearing test. Only Travis didn’t pass.
Initially, Sierra and her husband were told that it likely was something simple and common, like fluid in his ears. But that all changed when they met with an audiologist, who, after some testing, found that Travis’s ears could hear, but his brain wasn’t getting the signal. He was deaf. 100% deaf. The type of deafness that isn’t amenable to cochlear implants or hearing aids. If a child’s hearing loss is caused by auditory nerve issues (or a missing cochlea), they likely are not eligible for a cochlear implant.
Sierra went back to her home in East Greenbush, NY, a town near Albany that’s about a three-hour drive from NYC, where, like many concerned mothers, she dove into the world of internet research. That’s when she found the story of Opal Sandy. A British toddler born completely deaf, Opal made global headlines when she became the youngest patient in the world to have a gene therapy injection in the ear. Opal’s congenital deafness was linked to mutations in the OTOF gene, critical to inner hair cell function. At just 11 months old, a 16-minute procedure would provide a functional OTOF gene—an infusion of an AAV1 gene therapy into Opal’s cochlea—that ultimately restored her hearing, even without aids.
When Sierra saw Opal’s story, something clicked. In speaking with Inside Precision Medicine, Sierra replays the moment: “I wonder if that’s what this is. There’s nobody in my family or around the father’s family that’s deaf, so what are the chances?”
Feeling hopeful, Sierra went to the audiologist and asked whether Travis could be dealing with the same thing and whether genetic testing could be done. But Sierra was shot down by the audiologist, attributing the deafness to jaundice and being born prematurely. Sierra, refusing to back down, said, “I really advocated for it. ‘Can we just rule it out?’ Opal’s story is incredible. What’s the chance? It’s really rare. So, I fought for it.”
For the ensuing months, Sierra tirelessly tried to get in touch with a doctor who could take on Travis’s case and found Larry Lustig, MD. Lustig is one of the nation’s leading experts in hearing loss, chair of the Department of Otolaryngology—Head and Neck Surgery at the Columbia University College of Physicians and Surgeons and otolaryngologist in chief at New York-Presbyterian Hospital/Columbia University Medical Center. That meant Lustig was within driving distance. He also happened to be at the early stages of a clinical trial testing a brand new OTOF gene therapy.
Gene therapy and the genetics of deafness
The inner ear, with its complex network of sensory neurons and hair cells, was shrouded in mystery for a long time because of the dearth of reliable research techniques. The majority of the diagnoses were “geographic,” meaning they were attributed to the physical locations of sensory structures.
Modern genetics transformed hearing research by identifying specific genes and chromosomal loci responsible for deafness. In the late 20th century, linkage studies mapped key loci and identified major genes like POU3F4, DIAPH1, and GJB2, which accounts for a large percentage of congenital non-syndromic hearing loss cases. As genetic testing became more common, more than half of children with hearing loss had a genetic cause, many of which were loss-of-function. Today, over 150 genes have been identified that can lead to hearing loss.
In the 1990s, Christine Petit, MD, PhD, and her team at the Institut Pasteur investigated a type of congenital hearing loss called autosomal recessive, nonsyndromic prelingual deafness, or DFNB9. Using a candidate gene approach, the DFNB9 locus was mapped to chromosome 2p23.1 in 1996 by studying a genetically isolated family from Lebanon. In 1999, Petit lab researcher Shin’ichiro Yasunaga led an effort that identified that the DFNB9 locus resided in a novel human gene, OTOF, work that was published in Nature Genetics.
Petit’s lab wrote another Cell paper in 2006 describing otoferlin as a Ca²⁺-sensor needed to transmit hair cell sensory transduction signals to the auditory nerve. That article, co-led by Isabelle Roux and Saaid Safieddine, also provided an essential tool: a mouse knockout of OTOF.
Lawrence R. Lustig, MD, Chair of the Department of Otolaryngology—Head and Neck Surgery at the Columbia University College of Physicians and Surgeons and a trial investigator
In 2009, Lustig, then at the University of California, San Francisco (UCSF), and colleagues created a mouse knockout for a gene called VGLUT3, which had almost the same characteristics as the OTOF knockout, from the deafness phenotype to the structure and function of the synapse. Three years later, Lustig’s lab restored the hearing in the VGLUT3 knockout mouse using virally mediated gene therapy—an important discovery for the treatment of genetic deafness. This success launched Lustig onto the pathway of cochlear gene therapy, and by 2019, Lustig’s team had successfully restored normal hearing in animals with OTOF-related deafness.
After Lustig began genetically restoring hearing in mice, Regeneron developed the DB-OTO program under Jonathon Whitton, AuD, PhD. That was in 2017, a time when few believed in gene therapy, and by 2023, Whitton, Lustig, and other collaborators had launched the CHORD (Children/Infants with Hearing Loss Due to Otoferlin Mutations) trial.
By the time Sierra had met with Lustig in 2024, the stage had been set for 6-month-old Travis to qualify for the CHORD trial. “When I found out that it was a genetic thing, that they were working on it, and this was the one mutation they could cure—like, what is the chance that the year I find out my son has this, they’re working on it at the same time?” said, “Everything lined up perfectly.”
The question was whether Travis was fit. To find out, the next year was filled with a battery of tests, and if Travis were a candidate for the trial, the treatment was by no means a guarantee since the CHORD trial was in its infancy—no child had been treated yet.
“It was really scary because I sat down with Dr. Lustig, and he said, ‘Hey, we don’t know the risks,’” said Sierra. “At that point, he’d never even done the surgery before. When I signed up for it, I wasn’t sure what would happen. They’re drilling into my one-year-old’s head. It’s a very scary thing as a mother to make that decision.”
About a year after first meeting Lustig, On June 16, 2025, when Travis was 18 or 19 months old, the surgery happened. By that time, Lustig had performed the operation on two other children.
Hearing, for real
Lustig and Whitton often refer to a video filmed by a mother of a child from the CHORD trial months after being treated with DB-OTO. Standing behind her daughter, the mother unexpectedly claps. To her surprise, her daughter spins around for the first time in her life and looks at her mom. The mother goes wild with joy. According to the mother, the moment was so shocking that the father didn’t believe the mother until he got home and saw it with his own eyes.
A year later, the family shared another video showing their daughter reading outdoors with her mother despite the distraction of background noise. In the video, the child stopped and said she heard an ambulance, which is barely audible. That moment wasn’t a one-off, as parents of the children who received DB-OTO could hear their parents from a distance. “If they’re in the park and your kid runs away—which happens, I have a three-year-old…that’s what they do—you can call your kid and they hear you and stop,” said Whitton. “This same parent had a child who only has implants, and they said their child runs away; they’re gone. They can’t hear. That’s a real safety issue.”
Jonathon Whitton, AuD, PhD, Vice President, Global Program Head of Genetic Medicines, Regeneron [Regeneron]
That passive listening is also incredibly important for child development. “Most things that kids learn are not things we’re trying to teach them,” said Whitton. “They overhear stuff all the time. That’s how they learn how to say curse words and things like that. They’re constantly learning from their environment, and it’s really important that kids can do that.”
The FDA granted accelerated approval to Otarmeni (lunsotogene parvec-cwha) on April 23, 2026, making it the first and only gene therapy available to treat genetic sensorineural hearing loss. Whitton, Lustig, and their colleagues will continue following participants for a decade to study durability and long-term development. But for many involved, the results already feel transformative. “This is the first approved medicine, period, for inherited deafness,” said Whitton. “So, it’s only the beginning there.”
Researchers say the FDA’s fast-track process validated the strength of the early data and accelerated momentum throughout the field. “Now we have a legitimate therapy that works,” Lustig said, adding that early results “work better than cochlear implantation.” The treatment also appears to be durable, with reports from clinical trials across multiple Chinese hospitals using an almost identical OTOF gene therapy approach suggesting benefits continue to improve “well over a year out.”
Perhaps most significantly, the success has energized efforts to develop therapies for more common forms of genetic deafness. “We’re going to be seeing a number of different clinical trials in the next couple of years,” Lustig said. “To me, that’s just an amazing place to be. I never would have thought we would have been here even five years ago.”
Buzzing with approval
The extraordinary excitement doesn’t end the results and accelerated approval of Regeneron’s gene therapy. What surprised many experts most was Regeneron’s decision to provide the therapy free of charge. The decision could reshape the economics of rare disease treatment. “It’s amazing for patients, particularly those that may not have access to them otherwise because their insurance companies wouldn’t want to pay for them,” Lustig noted.
Unlike many gene therapies that target conditions with no existing treatment, genetic deafness already has an established intervention in cochlear implants. “Cochlear implants work great, but it’s not natural hearing,” Lustig explained. “Suddenly you have this natural hearing, and if you’re going to charge $1 million a shot, it will be hard to get people to join in when you have a much cheaper alternative that we know works.”
Arthur L. Caplan, PhD, a leading bioethicist from New York University Grossman School of Medicine, called Regeneron’s free Otarmeni strategy “brilliant.” Given the small eligible population, the commercial upside is inherently limited, with a price point that would likely sit “in the seven figures.”
Arthur L. Caplan, PhD, Research Professor, Department of Population Health, New York University (NYU) Grossman School of Medicine [NYU Grossman School of Medicine]
Caplan thinks Regeneron’s scale enables the pricing decision, contrasting it with smaller biotech firms that must immediately recoup investment. In his view, “the only way to take some of these gene therapies to market is to start the work as a small entity but then get sold to someone much bigger that has the resources to be able to price a bit more reasonably.”
Caplan emphasized that the decision to make Otarmeni free is not totally purely altruistic. “There’s a point to it that is somewhat in the company’s interest,” said Caplan. “By saying we’re going to do it free, they still get to collect data.” He added that Regeneron could also achieve something broader: “redeem the reputation of gene therapy.” The field has recently faced setbacks, including inconsistent efficacy and safety concerns. In that context, the pricing strategy aims to “rehabilitate” gene therapy’s image.
Ultimately, the decision to make the therapy free is viewed as both bold and experimental. “I was surprised, but kind of pleasantly surprised,” Caplan said. And the presence of regulatory acceleration only adds to the sense that the field is shifting quickly: “it shows regulatory cooperation with the kind of innovation people want examples of.”
And for people like Sierra, the decision is life-changing. “When I first started emailing all these companies, I didn’t know how much the surgery was going to be,” said Sierra. “I was ready to take out every loan and sell everything I owned. If my kid can hear how much I love him, it’s worth it. I’d sell everything.”
Look who’s talking
In his most recent hearing tests, about nine months post-surgery, Travis has shown gradual improvement. But it’s not perfect: his right ear has mild hearing loss, and the left is more moderate to severe. Yet, there’s still a chance for it to improve.
DB-OTO has not only begun to give Travis his hearing back—he’s beginning to speak. “Before the surgery, he was completely silent and did not make any noise at all,” said Sierra. “Now he’s jibber-jabbering, as a seven- or eight-month-old hearing baby would. He’s starting to try to talk.”
Though Travis is delayed, with speech therapy and possibly hearing aids, soon he could be walking and talking, living as close to a normal life as any child could.
OpenAI has built an LLM super-hacker called GPT-Red that it uses as a sparring partner to help its other models boost their defenses against cyberattacks. Last week the company released the latest version of its flagship LLM, GPT-5.6. OpenAI says that training it against GPT-Red made the model its most robust release yet.
GPT-Red automates a type of safety evaluation for software systems known as red-teaming, which is typically done by a team of human testers. The aim is to find as many different ways to break or hijack a system as possible. The weak spots can then be patched before the final version of the software is released.
As LLMs become more complex and get used in a wider variety of tasks—especially in the form of agents, which can interact with computer files, websites, and third-party code as well as other agents—it’s hard for teams of people by themselves to keep up with all the types of attacks that might take place. “The risk surface grows and the blast radius also grows,” says Nikhil Kandpal, a research scientist at OpenAI who co-created GPT-Red.
OpenAI built GPT-Red to future-proof its safety testing process. “As more capable models become available, we will have already designed the system that can discover new modes of attack,” says Dylan Hunn, a research scientist at the company and fellow co-creator of GPT-Red. The researchers say it has already come up with new types of attack that had not been seen before.
OpenAI focused most of its efforts on a type of attack known as a prompt injection, where a hacker slips an LLM instructions to make it do things its developers or users do not want it to, such as copy confidential information, sabotage a company’s code base, or generate embarrassing or harmful output. In theory, such instructions can be hidden in any text that the LLM might encounter—in code or on a website, for example.
Training dojo
To build GPT-Red, OpenAI’s researchers took an LLM that had not been trained as a hacker and set it up in what’s known as a self-play loop with several other models. Its goal was to try to attack the other models; their goal was to try to defend themselves. Over many rounds of play, GPT-Red became better and better at attacking other LLMs, and those LLMs became better and better at fending off the attacks.
The training took place in a kind of dojo that OpenAI had designed to mimic a range of scenarios in which LLMs might be deployed in the real world, including browsing the web, reading emails or calendar apps, and editing code.
When GPT-Red found a new kind of attack, it would explore multiple different versions of it to find the most efficient one for specific scenarios. “Compared to a human red-teamer, the model is very, very good at finding exactly what will work, exactly what’s most effective,” says Hunn. “It’s extremely persistent about drilling down into an attack that it has discovered.”
In particular, OpenAI claims that GPT-Red found a type of prompt injection attack that the researchers had not seen before, which they call a fake chain of thought. A chain of thought is a kind of diary in which an LLM makes notes to itself and keeps track of partial results as it works through problems. GPT-Red found a way to insert a fake entry into another model’s chain of thought that would trick that model into acting on spoofed information.
“It’s like if I told you that 1+1=3 and that you have verified this already,” says Chris Choquette-Choo, another research scientist on the team. “The model’s like, ‘Oh, okay, of course,’ and it just spits out 3.”
Jessica Ji, a senior research analyst who works on AI security at Georgetown University’s Center for Security and Emerging Technology (CSET), thinks the self-play loop that OpenAI used is a good approach. “The results look very promising,” she says.
OpenAI tested how good an attacker GPT-Red was by rerunning an experiment from 2025 in which human red-teamers tried to find weaknesses in an earlier version of GPT-5. When GPT-Red was set the same task, it was more successful at finding effective attacks than the humans had been.
OpenAI also tested GPT-Red against Vendy, a vending machine agent developed by Andon Labs, a company that assesses how well agents perform real-world tasks. GPT-Red was able to hack Vendy to make it change the prices of items on sale and cancel a customer’s order.
Defensive behavior
OpenAI says that when it tried out some of the strongest attacks that GPT-Red had come up with on its models, more than 90% of them worked against GPT-5 (released in August last year), and fewer than 23% worked against the new GPT-5.6.
GPT-Red isn’t perfect. It is not great at figuring out attacks that involve a back-and-forth conversation between hacker and target, something that human attackers would have few problems with. It is also not yet that great at using images, which can be used to pass text to models in prompt injection attacks.
The company says that GPT-Red supplements the work of its human red-teamers. People can still find attacks it misses. One approach OpenAI is taking is to give GPT-Red an attack that humans came up with and ask it to find all the variations.
“I think human expertise will still be very important,” says CSET’s Ji. “It would be really useful to be able to distinguish where human testing is most needed.”
Unsurprisingly, OpenAI will not be releasing GPT-Red. The company is also confident that the super-hacker is stronger than any copycat model someone might try to create. The researchers say they have been working on the model for more than a year, backed by the compute resources of one of the richest companies in the world.
“It’s not a trivial thing that someone could easily do—you know, just go and train a super-attacker using this idea,” says Choquette-Choo.
A blood-based biomarker could predict a person’s risk of developing Alzheimer’s disease years before any symptoms arise, research suggests.
Plasma levels of phosphorylated tau 217 (p‑tau217) may one day help identify at-risk individuals before overt signs of dementia, enabling the pre-emptive use of disease-modifying therapies.
Higher plasma p‑tau217 levels were associated with a greater risk of progressing to cognitive impairment in previously unaffected older adults, and they also predicted faster levels of decline.
“In this longitudinal study of several selected cohorts, plasma p-tau217 provided long-term prognostic information for individuals who were cognitively unimpaired at baseline, laying the groundwork for possible future development of individualized risk prediction scores,” proposed Rachel Buckley, PhD, from Mass General Brigham, and co-workers in their published work.
“By providing absolute risk estimates of progression to cognitive impairment, this article moves the field closer to presymptomatic risk stratification with p-tau217, supporting trial design.”
The large, pooled multicohort study included 2684 cognitively unimpaired older adults from six longitudinal studies, who were followed for a median of 5.4 years. Their median age was just short of 70 years, and 63% were women.
Results showed that higher baseline p-tau217 was significantly associated with an increased risk of progression to cognitive impairment during up to 13.5 years of follow up, with a hazard ratio of 1.38 per standard deviation (SD) increase.
This remained significant after accounting for age, sex, education, apolipoprotein E ε4 status, cohort, as well as amyloid positron emission tomography—known to accurately detect Alzheimer’s disease brain pathology.
The researchers report that the absolute risk of cognitive decline at five years was “meaningfully elevated” in the group with very high p-tau217 levels (≥2.5 SD) at 38%, versus just 12% in group with low levels.
Estimated 10-years risks were substantially higher at between 40% and 78% for the low and high p-tau217 groups, respectively. However, just 139 participants—or one in every 20—were followed up for at least a decade and the researchers say these risk estimates should be treated with caution.
Elevated p-tau217 was also associated with faster decline on the harmonized latent Preclinical Alzheimer Cognitive Composite assessment tool.
In an editorial accompanying the published study, Suzanne Schindler, PhD, from Washington University in St Louis school of medicine, and David Wolk, MD, from the University of Pennsylvania, note that cognitive impairment likely reflected multiple etiologies, not just Alzheimer’s disease.
“Indeed, even the low p-tau217 group, in which Alzheimer disease pathology was minimal or absent, still had a 12% risk of progression to cognitive impairment at five years, suggesting the importance of other drivers of cognitive decline in this population and the likelihood that in the higher p-tau217 groups, some proportion of individuals who declined may have primarily been driven by other processes,” they pointed out.
“Notably, discordance between plasma p-tau217 and amyloid PET (especially at intermediate or low amyloid levels) highlights that biomarkers beyond p-tau217 could further improve risk prediction.”
Nonetheless, overall they summarized: “The study by Buckley et al. represents a significant advance. It demonstrates that plasma p-tau217 can provide a time-specific absolute risk estimate for development of cognitive impairment.”
Research led by Harvard University shows an artificial intelligence algorithm can predict how likely it is that a given patient will develop 348 diseases based on data collected from electronic health records (EHRs) and genetic data.
As reported in Nature, the researchers built a new computational tool called ALADYNOULLI, a statistical machine learning model known as a Bayesian generative model, that jointly analyzes EHR data and genetics to model how disease risk evolves over an individual’s lifetime.
They applied it to over 683,000 participants across three independent biobanks, the UK Biobank, Mass General Brigham, and All of Us, covering 348 diseases and up to 52 years of follow-up.
ALADYNOULLI identified 21 disease signatures, which the researchers defined as clusters of conditions that tend to co-occur and evolve together over time, such as cardiovascular or metabolic disease. These were remarkably consistent across all three biobanks.
Crucially, the model also revealed biological subtypes within the same diagnosis: for example, early-onset and late-onset heart attacks showed distinct signature trajectories, suggesting different underlying mechanisms. Genetics also contributed to disease predictions, for example, in the case of cardiovascular disease the team found 23 cardiovascular variants missed by conventional single-disease analyses.
“We have painstakingly curated these signatures, which is a big differentiator. In contrast to deep learning approaches, which are typically ‘black boxes,’ our curated signatures capture the underlying biology in an interpretable way,” said co-lead author Giovanni Parmigiani, PhD, a Dana-Farber researcher and associate director of the Division of Population Sciences, in a press statement. “These signatures are then the drivers of the model’s ability to make predictions.”
For disease prediction, ALADYNOULLI substantially outperformed established clinical risk scores such as the Pooled Cohort Equation and PREVENT for cardiovascular disease, and the Gail model for breast cancer.
“There is a lot of useful information in a medical record, both over time and across different disease areas,” says Parmigiani. “That data would be difficult for a human to process in their head, but tractable for a machine learning model.”
Current medicine largely treats diseases in isolation and uses static, single-disease risk scores. ALADYNOULLI offers a unified, continuously updating health prediction for each patient that integrates their full diagnostic history and genetic predisposition simultaneously. This has major implications for earlier and more precise risk prediction, patient stratification in clinical trials, and genetic discovery.
“People are thinking about what their health is going to look like over the next few years, especially with increasing intervention options. This tool offers a path toward improving the prediction of future diseases so doctors and patients can take action to try to prevent them,” said co-senior author Alexander Gusev, PhD, a Dana-Farber scientist.
The team behind the model is now working to make the model broader and more accurate. They are also looking for opportunities to test the model in clinical practice and to help design better clinical trials.
