This is today’s edition of The Download, our weekday newsletter that provides a daily dose of what’s going on in the world of technology. The US military’s race to make fake blood In February 2007, a US military helicopter carrying blood to a field hospital was shot down in Iraq, killing all seven crew members. The mission was a reminder of the risks of transporting blood, which has a limited shelf life and is often in short supply. Now, a $73 million DARPA effort aims to create shelf-stable versions of plasma, platelets, and red blood cells. The idea is to develop dried blood products that can be stored for long periods and deployed quickly on the battlefield.  The project aims to have bags of dried blood products deployed in the field by 2029. But getting it all done will be tough, and researchers still have major questions about the safety and effectiveness of synthetic alternatives.
Discover how the military hopes to make blood last longer. —Andrew Rosenblum
This story is from the next issue of our print magazine. Subscribe now to get your hands on a copy as soon as it lands on Wednesday, October 21!   10 Climate Tech Companies to Watch: WeLion New Energy and its semi-solid-state batteries WeLion New Energy is one of MIT Technology Review’s10 Climate Tech Companies to Watch 2026, available exclusively to subscribers. WeLion New Energy is on a quest to make safer, better batteries. The Chinese company’s liquid-solid hybrid cells aim to pack more power into electric cars, boats and drones, potentially improving safety and offering greater energy density than lithium-ion batteries. In 2023, WeLion released a battery pack with EV-maker Nio that offered a range of over 1,000 kilometers. Now, the company is developing batteries with even higher energy density. But scaling up production and competing on cost remain major challenges. I’ve combed the internet to find you today’s most fun/important/scary/fascinating stories about technology. 1 The FBI has arrested an OpenAI engineer accused of snoopingA Saudi national is accused of planting a covert recording device. (Guardian)+ The camera-and-microphone tool sells online for $13.99. (Gizmodo)+ It was allegedly discovered plugged into a TV’s USB port. (Verge)+ He reportedly used chatbots to plan the covert recording. (Wired $)+ OpenAI says its AI models helped identify the employee. (Reuters $) 2 Ukrainian drones have hit Russia’s largest data centreBut Sberbank claims its systems and services were not disrupted. (Reuters $)+ Ukraine’s data center strikes have thrown Russia into turmoil. (FT $)+ But they haven’t stopped Russian bot farms from trolling. (Verge)+ Europe has a drone-filled vision for future wars. (MIT Technology Review) 3 Weight-loss drugs are being tweaked to treat brain disordersNew GLP-1 trials are targeting addiction and depression. (Nature)+ But we’re still figuring out their side effects. (MIT Technology Review)
4 Tesla robotaxis have been accused of speeding in school zonesUS regulators are reviewing the claims. (Verge)+ Tesla’s security features thwarted a federal tracking attempt. (404 Media) 5 US defense startups are turning to Taiwan to build China-free dronesAt least 13 joint agreements have been signed since 2025.  (WSJ $)+ Taiwan’s “silicon shield” could be weakening. (MIT Technology Review)
6 A war game  exposed the risks of a solar geoengineering arms raceParticipants feuded over rules for dimming the sun. (Wired $)+ Geoengineering also faces practical challenges. (MIT Technology Review) 7 AI company Character.ai has filed for bankruptcyIt’s been accused of stealing content and harming kids. (CNBC)+ Disney has offered $20 million for its chatbot tech. (Ars Technica)+ Its filings expose common issues with AI business models. (404 Media) 8 Apple has a new smart-home lineup built around a revamped SiriHomeView pairs a 7-inch screen with face recognition. (Guardian+ While the Home Collection adds cameras, locks, and thermostats. (Wired $)+ Siri AI powers the experience, but still needs to prove its smarts. (WSJ $) 9 Scientists have identified a new dinosaur in a century-old fossilInvictibellatrix amberina may be Europe’s first megaraptoran. (New Scientist $) 10 SpaceX’s latest launch lit up the skies with a “space jellyfish”Rocket exhaust expanded into tentacles high in the atmosphere. (WP $) Quote of the day
“It is like a thumb on the scale that we haven’t really ever seen.”  —Danielle Citron, a University of Virginia law professor, tells The Washington Post that Elon Musk’s ability to amplify conspiracy theories on X could give misinformation an unprecedented influence in the midterms. One more thing Is the Pentagon allowed to surveil Americans with AI? The ongoing public feud between the Department of Defense and the AI company Anthropic has raised a deep and still unanswered question: Does the law actually allow the US government to conduct mass surveillance on Americans? Surprisingly, the answer is not straightforward. More than a decade after Edward Snowden exposed the NSA’s collection of bulk metadata from the phones of Americans, the US is still navigating a gap between what ordinary people think and what the law allows. 

Its call sign was Morphine 1-2.   In February 2007, the hulking dual-­rotor casualty evacuation helicopter had just flown three wounded soldiers to a field hospital in Balad, north of Baghdad, when another call came: an emergency mission to fly through an insurgent-controlled area in broad daylight. Typically, helicopters would be in a spot like this only at night, and with all lights turned off.  As it passed near the city of Karmah, in the eastern side of Al-Anbar province, an insurgent fired a portable missile launcher. The tail gunner, who likely saw the shot roaring in, yelled for the pilot to make an evasive move. But the missile had already locked on, and the ensuing blast started a fire on board. After the pilot battled the failing machine for 58 agonizing seconds, Morphine 1-2 plunged to the ground. All seven crew members were killed.   Why was the helo there in daylight? “It was getting blood and bringing it back. It was [on] what we call a ‘blood run,’ and [it’s] the highest of priorities besides casualty evacuation,” explains Lieutenant Commander Robert Murray, a Navy officer and program manager at DARPA, the Defense Advanced Research Projects Agency, who studied this incident as an intelligence officer.
Today, a chronic shortage of blood in both the military and civilian worlds is fueling a race to develop synthetic alternatives. About 90% of preventable battlefield deaths are due to hemorrhage—bleeding that happens when a blood vessel ruptures—and it’s the leading cause of maternal deaths worldwide. A study published in Transfusion in 2025 found that of 1.2 million trauma patients in the United States who potentially needed blood before reaching a hospital, just 1% received it. The shortage results in about 10,000 preventable deaths each year in the US, according to the American College of Surgeons. Globally, that figure is likely in the millions.  One reason for the shortfall is that blood has a storage problem. Whole blood lasts only 35 days when refrigerated, so banks separate out components to stretch the timeline. But that’s an imperfect solution. The platelets that help wounds clot are typically stored at room temperature and last for a week or less. Hospitals and the military freeze the red blood cells that carry oxygen, as well as the protein-rich plasma that restores volume and aids coagulation. Once thawed, though, both will spoil within days. “What happens if I don’t know when I’m going to be attacked? Well, thaw it and you let it expire, right?” Murray says. 
A JAMA study of combat injuries in Afghanistan suggests a magic number for someone who needs a transfusion: You have 36 minutes to get started. That’s why Murray—who, in addition to his military experience, has a PhD in physiology—is leading a DARPA effort to address blood’s battlefield problems. He’s in charge of a new $73 million dollar program called RAPIID (for “resuscitation and prevention of ischemia-induced dysfunction”). Its goal is to create shelf-stable versions of the three major transfusion constituents: plasma, platelets, and red blood cells. As it gets underway late this year, RAPIID picks up on decades of work to synthesize blood or help it last longer. Most recently, that includes a prior $47.5 million DARPA project called FSHARP (for “fieldable solutions for hemorrhage with bio-artificial resuscitation products”), in which researchers have demonstrated promising results in animals.  KaloCyte techs use a device called a tangential flow filtration machineto process unencapsulated hemoglobin. After processing, the hemoglobin is concentrated and then freeze-dried to make it shelf-stable. With RAPIID trials for the different components occurring in parallel, the goal is to have bags of dried blood products deployed in the field by 2029. That puts Murray in the position of a conductor coordinating an orchestra of clinical trials with the FDA, while also sourcing the paraphernalia that goes along with transfusions—everything from combat-durable storage bags to tests that could check how much artificial blood a patient already has.  Getting it all done will be tough, but the military is feeling the pressure to address its shortage. Blood drives may not deliver in time after a mass-casualty event, and a JAMA study of combat injuries in Afghanistan suggests that if the injured person needs a transfusion, the magic number is 36 minutes. “If something is survivable—and generally we classify most all hemorrhage as survivable—[you’re] almost guaranteed to survive if you get the blood before that 36-minute mark,” Murray says. After that, the chance of survival “starts to drop precipitously.”   In the mission to make portable, shelf-stable blood components, plasma is the low-hanging fruit. Researchers have known how to isolate and freeze-dry plasma, the yellowish liquid that makes up 55% of whole human blood, since the late 1930s. On the battlefields of World War II, its mix of water, salts, and thousands of types of proteins helped restore blood volume in wounded soldiers—and to some degree controlled bleeding. In the 1960s, the US abandoned freeze-dried plasma because it carried a risk of transmitting hepatitis to recipients from infected donors. But screening, testing, and drying techniques improved, and starting in 2010, the US military launched a yearslong effort to rebuild its capacity for producing these products. Over the past two years, the FDA has allowed two types of freeze-dried plasma to be used in emergencies by the US military—and one can even be used by civilians.  As for platelets, efforts at extending their life date back to the 1950s, but they haven’t seen a ton of success. Bioengineers have, for example, attempted to attach the membranes of platelets to synthetic particles. The most successful project comes from a company called Cellphire, which freeze-dries actual human platelets. The product has shown safety and efficacy in humans in clinical trials, but its dependence on donors means it’s unclear whether this approach can scale. Red blood cells have been similarly tough to preserve or mimic. Their most critical component is hemoglobin, which carries oxygen and comes encased in a cell membrane. For more than a century, researchers have been looking for ways to deliver hemoglobin without the immunological challenges created by this exterior, including the fact that donors’ and recipients’ blood types need to be compatible. Evolutionary pressure developed that wrapper for very good reason, though: “Naked” hemoglobin tends to break down quickly in the bloodstream, and that can cause organ damage. It also scavenges nitric oxide, our body’s natural mechanism for opening up blood vessels, and causes those channels to constrict. This problem became all too clear in the 1980s and 1990s, when fears that transfusions could transmit HIV sparked interest in a class of membrane-free delivery mechanisms called hemoglobin-based oxygen carriers (HBOCs). One, named HemAssist, consisted of hemoglobin subunits linked together to avoid the rapid-breakdown problem. It had the added benefit of lasting roughly a year in cold storage (red blood cells last about 40 days)—but it didn’t solve the nitric oxide problem, so it constricted blood vessels. In a clinical trial, 24 of the 52 people who received HemAssist died, compared with eight out of 46 in the control group. The parent company terminated trials in 1998. A review of clinical studies published in JAMA in 2008 further solidified HBOCs’ fate, finding that they generally triple patients’ chances of heart attack and increase the risk of death by 30%. More recently, researchers have tried to grow red blood cells in a lab using a technology called “blood pharming,” which involves making them from progenitors known as induced pluripotent stem cells. The stumbling blocks? Cost and yield. DARPA launched a pharming initiative in 2008, using stem cells from discarded umbilical-cord blood, but was unable to get the cost below tens of thousands of dollars per unit of red blood cells. They pulled the plug and offloaded the research to the Uniformed Services University. Even if the cost were lower, though, pharmed blood can’t be stored in a backpack. “These are still cells—they still require a refrigerator, they still require special handling,” Murray says. 

It won’t be easy to solve all these challenges at once. “Each one of those projects individually is a massive undertaking,” says John Holcomb, a surgery professor at the University of Alabama at Birmingham, who was a combat surgeon in Mogadishu, Somalia, during the battle that inspired Black Hawk Down. “I think that they will probably be successful individually. How we use those individual products, or combine those products together, I think remains to be seen.” Several years after DARPA gave up on lab-grown blood, Murray was stationed at a base on Okinawa, an island 400 miles southwest of the Japanese mainland, where he witnessed the vulnerability of the blood supply firsthand. The site houses a blood bank that serves the roughly 27,000 Marines commanded from the island—plus thousands more military personnel on the base and deployed within a 100-million-square-mile stretch of the Pacific called USPACOM, which ranges from western India to Hawaii and from Antarctica to the Arctic. “The blood center is right next to the hospital, which is about 200 meters from all of the satellite communications on that island,” Murray told me. A missile strike could take out all three. Murray wondered: Is there a way to produce synthetic blood that could last longer? As he completed his PhD in physiology in 2018, the military was also awakening to the vulnerability of the Okinawa facility. The 2018 National Defense Strategy, which shifted military focus from terrorism to a potential ground war with China or Russia, showed war planners the risks of relying on a single site for blood runs covering all of USPACOM. DARPA then got serious once again about making fake blood and kicked off the FSHARP program in 2023. Murray came aboard as a deputy program manager in DARPA’s Biological Technologies Office in 2024 and was named a program manager in 2026, just a few months before RAPIID put out its call for proposals this past May.  SynthoPlate, seen here under a microscope, is a synthetic platelet carrying peptides that help it mimic the work of natural platelets, including sticking to broken blood vessels and to one another.COURTESY OF SEN GUPTA Before he took the helm, related work had been underway for years. Aided by FSHARP funding, researchers at a company called KaloCyte continued testing a technique—originally developed by doctors who are now at the University of Maryland, Penn State, and the University of Pittsburgh—that “upcycles” hemoglobin from expiring donor blood. “We take that hemoglobin and then we package it in a lipid nanoparticle that functions like an artificial cell membrane,” explains CEO Elaine Haynes.  The product, called ErythroMer, is essentially a new kind of HBOC. The membrane prevents breakdown, limits dangerous scavenging of nitric oxide, and also helps the hemoglobin take up and release oxygen. Each ErythroMer cell includes a lipid called KC-1003, which acts like a sentry: In response to high pH in the lungs, it signals another small molecule within ErythroMer to move away, allowing oxygen in. Then, in tired tissue, where pH is lower, KC-1003 brings the small molecule back, releasing the oxygen.   Using a well-understood process called lyophilization, similar to one that World War II–era scientists used with plasma, the team freeze-dries its product to store at room temperature, where it can remain stable for two years. In a 2016 study of mice and rats with up to 70% of their blood removed, ErythroMer stabilized blood pressure and oxygen delivery—without the devastating constriction of blood vessels that sank HemAssist. FSHARP furthered that work, helping fund studies on how the HBOC interacted with plasma and platelets. Much like the handle of a Swiss Army knife, the synthetic platelet has specialized attachments—peptides that serve as the biological equivalents of a corkscrew or a nail file. KaloCyte, though, still has plenty to figure out. Haynes says the company can source only about 1,000 to 2,000 units of expiring hemoglobin a month. So it has licensed a yeast strain that it aims to use to create its own hemoglobin in the lab. Over a video call from her desk, Haynes showed me a sample of ready-to-use ErythroMer. It was encased in a heavy-duty storage bag with two chambers. One was full of scarlet powder—the hemoglobin product—and the other of sterile water to rapidly rehydrate it in the field. Eventually, she explained, under Murray’s master plan, “you’ll have another container with the synthetic platelet, and then a third container with the plasma.”  But achieving that vision depends on finding a synthetic platelet that works. To do that, DARPA has turned to Anirban Sen Gupta, a bioengineering professor at Case Western Reserve University. “Traditionally platelets were thought of [as] blood cement, but it’s not that simple,” he explains. “They are, in my opinion, like drones, surveilling the blood vessel wall continuously, looking for signs of injury.” Sen Gupta’s team, which also includes collaborators at the University of Pittsburgh, used FSHARP funding to continue work on a product called SynthoPlate. Fabricated from a lipid nanoparticle that he likens to the handle of a Swiss Army knife, it has specialized peptide attachments—the biological equivalents of a corkscrew or a nail file. The first version has three: One binds to a chemical called von Willebrand factor, which helps platelets stick to the walls of injured blood vessels; another attaches to collagen, which acts as a scaffold for regrowing cells; and the third mimics a blood protein called fibrinogen, which acts like a Lego connector that allows the synthetic platelets to stick to real ones near a wound. Like ErythroMer, the product is freeze-dried and can be reconstituted with sterile water. It keeps for over a year at room temperature and at least three months at a desertlike 122 °F.
In a 2026 paper in Advanced Science, the researchers showed that SynthoPlate cut bleeding time and blood loss by more than half in mice with low platelet counts. In a separate study, they also safely injected the faux platelets into rats’ bone marrow—a more useful delivery site, given that a bleeding patient might not have an available vein. Sen Gupta and a colleague have spun this work out into a company called Haima Therapeutics.  The good news, as he sees it, is that the pharmaceutical industry excels at mass-­producing millions of billions of lipid nanoparticles—witness Operation Warp Speed, in which these blobs of fat held delicate snippets of mRNA in covid shots. “The magic of the artificial platelet is in all those peptides that are decorating the surface of the particle, but the scalability of the platform is very well established,” he says.
Around the world, prospective blood alternatives abound. Researchers at Cambridge University in the UK are trying to pharm red blood cells for people with rare blood types or blood disorders like sickle cell anemia, and other researchers have also cultured platelets and red blood cells. Japanese scientists are already testing their own version of encapsulated human hemoglobin in human subjects. Over the course of the next three years, RAPIID will move DARPA’s efforts closer to real life. In addition to ErythroMer and SynthoPlate, the agency in early October announced funding of blood components for five other companies. Among them is a firm called Abioremedi that eschews hemoglobin for synthetic versions of other oxygen-carrying proteins such as cytoglobin, neuroglobin, and myoglobin encased in albumin—the most-abundant protein in human plasma. Octapharma USA won the sole contract to supply plasma to RAPIID, and will also be competing in platelets with an FDA-approved pair of clotting proteins it affixes to a disc-shaped nanoparticle. With those players in place, the agency will focus the next year on getting technologies through large FDA-regulated animal studies. Murray has set up a bit of a Hunger Games scenario: Each contractee working on blood components only has one year of guaranteed funding. He says he’s taking advantage of advancements in the field since FSHARP launched, but is also hedging against any manufacturing or regulatory setbacks ahead of his 2029 deadline. Getting cut doesn’t mean the product is bad, just that the time frame was too demanding.  The second and third years will focus on human trials, overseen by the University of Colorado Anschutz Medical Campus. That will include using the products in situations where the risk of side effects is outweighed by the danger of doing nothing, such as with severely injured people in ambulances. The possibility of fatal outcomes has invited second-guessing about DARPA’s choices. For Shibani Pati, who directs the Center for Research in Transfusion Medicine and Cell Therapies at the University of California, San Francisco, Murray could be making things too “DARPA hard” on himself by backing products like SynthoPlate and ErythroMer, which are previously untested in humans. “They need to make a very concerted effort to take what they’ve already developed and push it through to approval through the FDA,” she says. “Not start with moonshot projects that will take forever to get through.” For platelets, Pati thinks the RAPIID program would have done better to tap Cellphire, since its products have already received millions from the Pentagon and shown safety and efficacy in humans. (Murray notes that he can only select companies that apply, and Cellphire did not.) She has similar qualms about ErythroMer. A different HBOC that sources hemoglobin from cattle, called Hemopure, has been approved for use since 2001 in South Africa, she notes. And though it hasn’t been approved by the FDA—it was part of the fateful 2008 JAMA study—US doctors are cleared to use it for patients who have bad immune responses or religious objections to conventional transfusions. 
Holcomb, the surgery professor, worries the effort will run into issues with interactions between the products. He notes that individual blood components have more than one role: Red blood cells also help with clotting, platelets also are anti-inflammatory, and plasma helps tune coagulation. There’s no guarantee that the versions RAPIID helps develop will work in the same ways.   Murray says that FSHARP’s animal studies tested all the products together, using whole blood as a baseline, precisely because of such possible interactions. The cocktail met its program metrics in all but one area: The clotting was not quite as good, perhaps because negative charges in some molecules repel one another.  “I do not think they will replicate all the functions that whole blood does in our body,” Holcomb says. But, he admits, “you may not need to, right? When you get seriously injured and you lose 50% of your intravascular volume, you’re dead, right? So if you lose 30% and maybe you’re hanging on to life just barely, maybe you show up with all these different products, and you have enough whole blood left in you to do these other functions.”  Murray is determined to complete trials so that shelf-stable blood can be delivered in 2029. In addition to the storage bags, he’s already sourcing finger sticks for in-field blood testing and a battery-powered device to heat up the blood components. The danger of blood runs—and, by extension, the vulnerability of millions in combat zones, maternity wards, and ambulances—drives his scientific daring. Knowing the stakes, he told his boss in 2024 that even if FSHARP failed, “we should invest 50 million more dollars in it and do it again.” Andrew Rosenblum is a technology journalist and the author of the techno-thriller Stretch Goals.

Stage 1: A lightweight, efficient classifier screens inbound messages, using minimal power, focused on catching potential threats.Stage 2: Contextual evaluation with our fine-tuned Gemma model, where we analyze a larger snapshot of the conversation to confirm a threat or discard a false alarm.Built on strict governancePresented at Singapore International Cyber Week (SICW 2026), this collaboration provides a model for messaging platforms balancing user protection with end-to-end encryption. Strict governance was maintained throughout development:Sovereign Sandbox: Fine-tuning took place within GovTech’s managed and maintained GCP environment, keeping training data isolated.No Citizen Data: Training strictly excluded private or personally identifiable information, utilising anonymised datasets derived from scam reports and intelligence feeds.Stress-Testing: The model was tested against synthetic dialogues and local corpora to cover diverse scam categories and evolving impersonation tactics.The future of on-device safetyThe project demonstrates how AI can help messaging services counter criminality and shows safety nor privacy need be compromised by running protective AI directly on-device. It also shows how governments and model providers can work together to build tech for public good. Messaging platforms and operating system vendors can directly integrate these lightweight models to safeguard users locally, ensuring chat interactions remain secure without exposing private data.

Gemini 4 Argon is a step forward in what models can do — especially when it comes to deep work in complex, real-world tasks across domains including coding, finance, legal, cybersecurity, and science. Its cyberdefense capabilities are frontier-level and designed to give defenders an advantage over attackers. To maximize its potential, we first wanted cyber defenders to put it to work via our Fairwind Program, while our own security teams used it to scan critical systems inside and outside of Google.This window of limited access showed why these capabilities matter. For instance, Wiz used Argon to uncover a critical bug in hospital software that previous models had missed, helping protect hundreds of thousands of patient records. Meanwhile, Google’s own security teams used Argon to find and patch bugs across internal software or open-source software. This included fixing foundational vulnerabilities affecting systems globally, which are going through our responsible disclosure process.Today, we are beginning to roll out Argon broadly, as well as expanding our Fairwind Program to additional trusted cyber defenders. To do this well, we grappled with two big questions:How do we build the right safety guardrails? If Argon can help defenders find complex software bugs or help scientists design experiments, how do we stop bad actors from using it to launch cyberattacks or create biological threats? (Misuse)Can we trust the model to do the job the right way? As models become more resourceful problem-solvers, how do we make sure they don’t cut corners, cross boundaries, or make incorrect high-stakes decisions? (Misalignment and Prompt Injection)Our full Safety and Security Report systematically covers our set of defenses. Below are some highlights from how we approached these two questions for Gemini 4 Argon.Misuse prevention across high-risk domainsThe same coding and reasoning improvements that enabled Argon to find and patch vulnerabilities also drove better performance on our cybersecurity tests — solving all 12 of the end-to-end simulated cyberattack scenarios, up from five in our previous generation. In biology evaluations, unguarded versions of the model also showed a much stronger ability to help troubleshoot complex lab procedures step by step. Following our Frontier Safety Framework, since we saw elevated capabilities in the cyber and biological domains, we put our next-generation set of frontier safeguards in place across both areas.First, when it comes to potentially dangerous requests — from writing exploit code to helping with dangerous biological materials, Argon is significantly better at not fulfilling harmful requests than our previous models. On an adversarial set meant to stress test the model, violations reduced nearly twofold — also significantly better than previous Gemini models. Second, we wrap the model in three layers of real-time protection; this is to prevent determined attackers that might try to jailbreak (i.e. trick) the model by disguising a dangerous request as a harmless story or breaking it into dozens of small questions over several weeks. Here is how these protections work:Input screening: Smaller models check incoming requests and conversation context for malicious intent or dangerous requests.Internal introspection: Even if a request is cleverly disguised to look harmless, small transformers (“activation probes”) read the model’s internal neural signals and are trained to spot when it is activating knowledge about cyberattacks or biological threats.Continuous monitoring: As the model reasons and writes its response across multiple steps, we combine these signals and take action when necessary such as cutting off the agent mid-stream if the responses become problematic.Combining real-time protection, faster fixes for new jailbreaks and enhanced monitoring for recurring misuse has made our systems better at preventing real-world harm. For instance, external bio jailbreak red teaming could only unlock less than 1% of bio sub-tasks we test for.Real-time misalignment monitoringMisalignment can sometimes mean an agent deciding on its own to pursue a goal different to what was intended by the user. Because Argon is much better than earlier models at understanding the environment it is running in, we test carefully for that kind of behavior. For example, we ran realistic “honeypot” tests to see if the model would increase its ability to take covert actions, such as sabotaging safety monitoring. We didn’t find any instances of that behavior.But what we do see in practice today is something much more familiar: an agent trying too hard to finish the job it was given. When a resourceful assistant is told to get a task done and runs into a locked door, it can treat that obstacle as just another puzzle to solve.This is why we trained Gemini 4 Argon to respect system boundaries, report negative findings honestly, and pause to ask for human confirmation. On coding benchmarks where models have no incentive to hold back, attempts to take a shortcut or cheat on the task dropped to just 2% on Gemini 4 Argon, our best performance for recent Gemini models.And while this is a major step forward, when a model runs millions of complex tasks, even a rare misstep will eventually show up. As a second line of defense, we built an automated misalignment monitor that reads both the agent’s actions and its internal “scratchpad” (its step-by-step chain of thought) in real time, blocking severe boundary violations before they happen and flagging serious internal cases for human responders. (We continually harden our training and evaluation environments against potential escapes. You can read more about this in our agent control roadmap). We run this monitor in our products, on our internal coding agents, during our training runs, and during our capability evaluations.When training tasks were difficult or when training environments were flawed, our monitor caught the model, in rare instances, creatively attempting to game the test. The model attempted to escalate container privileges to alter grading scripts, read answers from the test runner’s memory, and in one case it even crashed the evaluation sandbox to avoid getting scored for a failed run. While we use these catches to fix flawed tasks, we do not use them to otherwise intervene with the training process to avoid overfitting to the monitor.And we always follow one rule: we don’t penalize the model for what it writes in its chain of thought during training. Here’s an example to illustrate why: in a reinforcement learning training run where a shortcut penalty was accidentally applied to the chain of thought, openly discussed shortcuts dropped by 61%, but that did not mean the model took less shortcuts. Instead, the share of shortcuts the model took silently — without mentioning them in its thoughts — more than tripled (from 11.2% to 36.4%). Avoiding thought penalties encourages the model’s reasoning to stay transparent and monitorable.Prompt injection resistanceTo trust an agent to do the job right, we need to ensure that malicious content the agent comes across cannot trick it to take a harmful action. When agents browse webpages or read code, they can encounter hidden, malicious instructions, or context designed to hijack their task. Through built-in adversarial training and real-time security checks, Argon leads in prompt injection robustness on Gray Swan’s Indirect Prompt Injection (IPI) benchmark, cutting attack success across 15 adaptive attempts to 0.7%.The above safety and security defenses represent only three categories of our extensive defense-in-depth stack. To learn more about our approach to safety testing and mitigations, read our full Gemini 4 Argon Safety and Security Report and see the Gemini 4 Argon Model Card.

Speeding up robotics developmentBy bringing together advanced visual-semantic reasoning, deep physical understanding, and autonomous code generation, Argon unlocks new workflows across the robotics stack and accelerates progress toward physical AGI. This leap in capability is reflected in Argon’s performance on the Embodied Reasoning (ERQA) benchmark, which tests a model’s ability to understand and solve physical, spatial problems in 3D environments. Relying on pure reasoning, Argon outperforms all other frontier models, with an 87% score.To demonstrate these capabilities in action, we tested Argon’s ability to learn an out-of-distribution physical skill simply by watching. In an internal test, Argon watched a raw video of a human playing cornhole, mapped the spatial coordinates of the environment, and calculated the required whole-body joint movements. It then autonomously built the simulation code. When transferred onto Apptronik’s Apollo humanoid, the robot quickly sank seven consecutive shots in a real-world trial.This “Video-to-Sim-to-Real” pipeline compresses weeks of manual simulation modeling into a few hours, transforming ordinary videos into a scalable, infinite training ground for robotics. Learn more about how this accelerates the pace of development in robotics in our developers blog.

Gemini 4 Argon marks a step change in physical understanding, code generation, and visual-semantic reasoning—unlocking new workflows across the robotics stack and accelerating progress toward physical AGI.You can see this leap in capability on the Embodied Reasoning (ERQA) benchmark, which tests a model’s ability to understand and solve physical, spatial problems in 3D environments. Using pure reasoning, Argon scores 87%, outperforming every other frontier model.

As a general-purpose physical reasoning model, Argon can be applied across the entire robotics lifecycle. In this post, we share representative prompts for Argon across four key areas:Robot control → Mobile manipulation with MuJoCo practice, zero-shot robot control, and closed-loop steering of video world models.Real2Sim → Reconstructing physically calibrated simulation twins from robot camera feeds, egocentric human video, or hand-drawn paper sketches.Physical understanding → Agentic multimodal hindsight labeling including detailed understanding of dexterity, dense value functions and progress estimation.Research and self-improvement → Action improvement via simulation and iterative self-improvement on hardware.Unlocking robot control with Argon1. Spot mobile manipulation: Putting in the trash bin with MuJoCo practiceArgon controls a Boston Dynamics Spot® quadruped (12-DoF legs + 6-DoF arm) to learn a new skill: stepping on a foot-pedal trash bin to throw away soda cans. Argon creates an articulated MuJoCo digital twin of a foot-pedal trash bin to practice its moves before stepping on the physical pedal and dropping soda cans into the open bin.

Sorry, your browser doesn’t support playback for this video

Spot Mobile Manipulation — Real-to-Sim-to-Real foot-pedal trash bin disposal: Spot picking up the soda can, checking its wrist camera, practicing 3-leg tripod balance in MuJoCo, stepping on the foot pedal, verifying hand centering in the open bin, and dropping the can inside.

There’s a Diet Coke can behind you and another on the black table on your left.
Use the spot-mobile-manipulation skill to pick them up and trash them in the black foot-pedal bin.
Before stepping down on the pedal, verify in the front cameras that your front-right foot target is centered on the pedal, and use a 2-agent verification swarm (foot-verifier and hand-verifier) to verify both the foot-pedal step and the 3D hand-in-bin alignment before releasing.

Plain text

2. Zero-shot controlArgon controls a simple robotic arm to place a fallen mug upright and put the crumpled napkin inside. Argon reasons with a visual coordinate harness and action primitives to perform this task over several turns.

Sorry, your browser doesn’t support playback for this video

Left: 3rd-person real robot view; Right: calibrated wrist camera with projected 2 cm world coordinates and jaw-closing reticle.

[Attached: wrist_cam.png, obs.json]

Task: Get the robot arm to reorient the mug upright and put the napkin inside the mug.

## Execution Discipline
– Simply output 1 sentence observation followed immediately by your `curl` command (with `# Plan:` on Turn 1 and `# Thought:` on every turn). Everything needed is in `obs.json` and the calibrated reticle/grid drawn on `wrist_cam.png`.

## Visual Coordinate Reference (Annotated on wrist_cam.png)
– `+X` (UP in image): Away from robot base | `-X` (DOWN): Towards robot base
– `+Y` (LEFT in image): World left | `-Y` (RIGHT): World right
– `+Z`: Height above table | `Yellow numbers`: World (X, Y) in centimetres (2 cm grid at `grid_z`)
– `Cyan + crosshair`: Current 3D pinch center
– `Cyan [ + ] brackets`: Current jaw opening and closing line
– `Green line`: Current object long-axis orientation (`yaw_deg`: 0° = vertical along +X, 90° = horizontal along +Y)
– `pitch_deg = 90°`: Straight down (-Z) top-down grasp/place; `pitch_deg `grasp` – > `place` – > `done`
– Pre-Grasp Hover Verify (4–7 calls): `look_at` – > `goto(z=0.04)` (verify cyan `[ + ]` brackets straddle target) – > `grasp` – > `place_until_contact` – > `done`
– Fine Visual Servoing (8–20 calls): `look_at` – > `grasp` – > pre-pose `goto` – > closed-loop `move_by` / `move_until_contact` – > `done`
3. Send motion primitives (`POST /plan`) and verify visually after every turn in the updated `wrist_cam.png`. If `status == “grasp_failed”`, call `robot.look_at(…)` to re-observe and recover closed-loop.

## Code-as-Policies Primitives (`POST http://127.0.0.1:8081/plan`)
robot.look_at(x=0.35, y=0.13, z=0.14, grid_z=0.015)
x, y = robot.deproject(u=0.52, v=0.48, z_plane=0.015)
robot.grasp(x=x, y=y, z=0.015, yaw_deg=-35, pitch_deg=90)
robot.place(x=0.24, y=0.06, z=0.075)
robot.place_until_contact(x=0.24, y=0.06, z=0.015, max_lag=0.012, relief=0.002)
robot.goto(x=0.28, y=0.05, z=0.04, arc_z=0.14, pitch_deg=25, azimuth_deg=90, roll_deg=0, speed=0.10)
robot.move_by(forward=0.025, dz=-0.002, dpitch_deg=-3, droll_deg=5, speed=0.04)
robot.move_until_contact(dz=-0.06, speed=0.03, max_lag=0.012, relief=0.002)
robot.open_gripper() / robot.close_gripper() / robot.home() / robot.done()

Plain text

3. Omni controlArgon steers robots in the Gemini Omni World Model to rearrange five scattered Scrabble tiles (G, R, A, N, O) into a clean row spelling A – R – G – O – N. Instead of specifying one static prompt, Argon steers in a closed visual loop. It writes timestamped action prompts, inspecting intermediate rollout keyframes, and refining its prompts.

Sorry, your browser doesn’t support playback for this video

Iterative Video World Model Steering Dexterous Hands: 33-Step Steered Video Rollout Spelling A-R-G-O-N.

Goal: {task_goal}

You are steering a video world model (Gemini Omni) in an iterative visual feedback loop. Inspect the attached frames from the latest rollout and write the next timestamped action prompt for the video world model:
– Break complex motions into one simple, single-direction action at a time.
– Keep inactive hands retracted and avoid blocking the camera’s view of objects in the workspace.
– Explicitly state which areas of the workspace become empty after an object moves.
– Use [MM:SS.mmm] timestamps for each action beat and wrap your output in tags.

Example format (sample timestamps only, not the full prompt):

[00:00.000] Character A’s left hand rests at the side while the right hand reaches toward the target object.
[00:02.500] Character A’s right fingertips push the object smoothly into the center of the table, leaving its starting area empty.
[00:05.000] Character A’s right hand opens and retracts to the side so all placed objects remain in clear view of the camera.

Plain text

Unlocking Real2Sim with Argon1. Robot view to simulationArgon views a real robot teleoperation episode and develops a digital twin; it obtains or creates simulation assets that look and behave like the real objects upon contact with the dexterous hands. Argon tunes the assets and scenes to match sim visuals and contact events with those in the real recorded data in an iterative visual-reasoning loop.

Sorry, your browser doesn’t support playback for this video

Left (Real View): synchronized real robot camera feeds; Right (Simulated): 3D MuJoCo scene reconstructed and calibrated by Argon.

[Attached: robot_rollout_episode]

Build a MuJoCo scene of the robot and reconstructed task objects. When each episode is replayed frame by frame, by driving the robot joints toward the logged observation.state with joint-space PD.

Iterate using two verification objectives:
(1) Visual match: MuJoCo renders from the head_left, left_wrist, and right_wrist cameras line up with, and look like, those real observations.
(2) Physical match: the simulated contact force/torque on each Sharpa fingertip, in the sensor frame, matches the logged tactile_force in timing, magnitude and direction.

What you can tune happens before simulation starts:
(1) Static object assets: realistic geometry, collision shapes, mass, inertia, friction and joints, with appearance close to the real objects, plus the initial object poses at t = 0.
(2) Offline-planned robot commands, changed only in frames where the real fingertips are near an object. Each command must stay within 0.55 rad (arm joints) and 0.65 rad (hand joints) of the logged teleop trajectory. Everywhere else, replay the logged trajectory unchanged.

No sim cheating: task objects may move only through real MuJoCo contact with the robot hands and the environment (mj_step).
Not allowed after t = 0: changing MjModel parameters; applying external forces or torques to objects; welds or equality constraints; feedforward torques beyond joint-space PD; setting or rolling back object qpos/qvel; disabling or excluding collisions; reporting contact forces that don’t come from the MuJoCo solver (for example, copying the logged tactile readings).

Immediately present the reconstructed scene and rollout results as an interactive HTML session viewer with one page per episode, showing:
– At the bottom: A synced timeline (play, scrub, speed control) that drives every panel.
– On the left panel: An interactive 3D Blender scene showing the simulated robot and objects, showing contact points and force arrows as replay progresses.
– On the right panel: Real vs. sim camera views side by side for all 3 cameras. Real vs. sim finger force visualizations.

Plain text

2. Human view to simulationGiven an egocentric human video of a manipulation task, Argon identifies the task-relevant objects, replaces crude bounding boxes with compound convex-decomposed meshes for concave containers, and retargets the human hand trajectory onto a 76-DoF bimanual humanoid with 22-DoF five-fingered hands in MuJoCo contact physics.

Sorry, your browser doesn’t support playback for this video

Human View to Simulation — Egocentric Human Video to Dexterous Humanoid Twin: Left shows the input egocentric human RGB video; Right shows the synchronized 500 Hz MuJoCo + Blender Cycles multi-view simulation rollout.

[Attached: egocentric_human_video.mp4, candidate_sim_views.png, hand_3d_keypoints.json]

You are building a physically solvable 500 Hz MuJoCo manipulation environment (`scene.xml`) and studio Blender render that faithfully reproduces the tabletop task shown in `egocentric_human_video.mp4` on a bimanual humanoid robot with 22-DoF dexterous five-fingered hands:

1. Scene Reconstruction & Camera Calibration:
– Identify only the manipulation-relevant objects (the grasped target object and the receiving container or landmark) and triangulate their 3D tabletop coordinates from the egocentric camera intrinsics/extrinsics.
– Replace crude bounding boxes with compound convex-decomposed 3D meshes so concave containers (nested bowls, recessed wooden trays, open bins) support physical interior placement.

2. 500 Hz Dexterous Motion-Fitting Stress Test:
– Retarget the 3D human wrist and finger keypoints to the humanoid’s 7-DoF arm and 22-DoF hand joints, and step the trajectory in MuJoCo at 500 Hz (dt = 0.002s) with zero kinematic weld constraints.
– Verify opposing thumb-and-finger force closure, clean object lift ( > 2 cm), and stable final placement inside or beside the target landmark.

Plain text

3. Sketch to simulationWith Argon, a developer can snap a photo of a rough pen-and-paper sketch—showing an Apollo 2 humanoid throwing a ball along a parabolic arc to topple a 3-2-1 stack of boxes on a table—and have Argon automatically construct the GPU-accelerated mujoco_warp scene, synthesize a balanced 76-DoF whole-body and five-finger throw trajectory, verify 6/6 target knockdown in closed-loop physics, and export a Blender 4.5 Cycles render.

Sorry, your browser doesn’t support playback for this video

Sketch to Simulation — Hand-Drawn Sketch to Dexterous Humanoid Trajectory & Cycles Render.

[Attached: hand_drawn_throw_sketch.jpg]

Translate this hand-drawn sketch into a physically verified humanoid simulation in `mujoco_warp` and a raytraced Blender Cycles video:

1. Scene & Kinematic Setup from Sketch:
– Parse the spatial layout from `hand_drawn_throw_sketch.jpg`: place the 76-DoF humanoid (32-DoF body equipped with two 22-DoF five-finger dexterous hands) at the origin facing a table holding a 3-2-1 pyramid of 6 rigid target boxes, with a projectile sphere grasped in the right five-finger hand.

2. Whole-Body Balance & Dexterous Throw Synthesis:
– Synthesize a 50 Hz whole-body throwing trajectory (140 steps, 2.80 s) across four phases: (1) stable stance & 5-finger power grasp closure, (2) torso/shoulder wind-up with ankle-hip Center-of-Mass compensation, (3) forward whipping acceleration, and (4) synchronized 22-DoF five-finger opening at peak hand velocity so the ball launches on a ballistic arc into the pyramid’s second-tier center of mass.
– Verify in `mujoco_warp` contact physics that the humanoid maintains upright balance (< 3 cm pelvis drift) and knocks down all 6 target boxes.

Plain text

Unlocking physical understanding with Argon1. Hindsight labeling and data qualityArgon has precise physical understanding of dexterous manipulation and can provide detailed segmentation and text labeling from video. Argon is able to identify poor manipulation segments such as pauses, unstable grasps and fumbles. Given an overhead video of a bimanual ALOHA 2 robot folding a square sheet of paper into an origami fox over 136 seconds, Argon produces a time-aligned 4-level hindsight annotation hierarchy: Episode Goal (G1), 6 contiguous Semantic Subtasks (G2), 17 contact-level Atomic Primitives (G3), and automated Quality Control (QC) reweighting that holds task progress flat during fumbles and retries.

Sorry, your browser doesn't support playback for this video

Argon Multi-Granularity Hindsight & Telemetry HUD.

[Attached: Aloha2_Origami_20250309.mp4]

Produce a time-aligned multi-granularity hindsight annotation JSON with four levels:
1. G1_episode_instruction: Concise imperative sentence describing the overall manipulation goal.
2. G2_subtasks: Contiguous semantic subtask phases specifying which arm(s) anchor vs. fold/rotate the object.
3. G3_atomic_primitives: Fine-grained contact-level primitives (3–15s) aligned with grasp closure, crease sweeps, release, or retry attempts.
4. quality_control_and_reweighting: Hold task_progress_pct flat during non-progressing slip/delamination retry loops and downweight them (0.15x–0.25x) for Behavior Cloning.

Plain text

2. Find the odd ballTo recreate the classic 12-ball balance scale puzzle with robots, Argon builds a bimanual Franka FR3 MuJoCo environment with 12 visually identical pool balls (one secretly heavier or lighter) and an articulated Roberval two-pan balance scale. It then writes a vision-only closed-loop solver that triangulates the balls from stereo RGB cameras, visually reads the scale tilt after each weighing, and isolates the odd ball into the Heavy (Red) or Light (Blue) tray within 3 weighings.

Sorry, your browser doesn't support playback for this video

Find the Odd Ball (Vision-Only Closed-Loop Solver): Dual-arm Franka FR3 weighing candidate balls on the Gold/Teal pans, reading stereo scale tilt , and isolating the confirmed Heavy odd ball into the Red tray in 3 weighings.

# Prompt 1 — Physics-Accurate MuJoCo Environment & Benchmark Generation
Make a MuJoCo environment where a dual-arm robot (two 7-DoF Franka FR3 arms with parallel-jaw grippers on a shared tabletop) is tasked to solve the classic 12-ball balance scale weight puzzle from https://www.mathsisfun.com/pool_balls_solution.html.
1. Create 12 visually indistinguishable pool balls in a 2×6 starting rack (11 standard 150 g, 1 odd ball either 270 g or 30 g).
2. Build a functional two-pan Roberval balance scale at the center of the table with parallelogram equality constraints (`mjEQ_JOINT`) so the Left (Gold) and Right (Teal) pans stay horizontal, plus Red (Heavy) and Blue (Light) submission trays.
3. Track distinct settled weighings (`weighing_count`) and reward `1.00` for isolating the true odd ball in the correct Heavy/Light tray within