A simple pattern of black-and-white stripes could cause an autonomous vehicle or robot to misjudge how far away an obstacle is, potentially triggering an unexpected maneuver or even a collision, according to new University of Florida research.

The finding exposes a previously overlooked vulnerability in the technology that autonomous systems use to perceive depth—a critical capability for vehicles, drones and robots navigating their surroundings. The study will be presented in November at the Association for Computing Machinery Conference on Computer and Communications Security, or ACM CCS and appears on the preprint server arXiv.

Researchers found that repeated patterns in the environment can interfere with the algorithms and artificial intelligence models that estimate the distance between a camera and an object. By manipulating the pattern, an attacker could make an obstacle appear significantly closer or farther away than it really is.

But the problem doesn’t necessarily require an attacker, said Sara Rampazzi, Ph.D., an assistant professor in the Department of Computer and Information Science and Engineering at UF.

“These things can also happen naturally, so it’s not a matter of imagining a sophisticated attacker. It becomes a safety problem,” said Rampazzi, lead of the research effort.

Repeated patterns distort stereo depth

The research focuses on autonomous systems that use stereo cameras—two cameras positioned close together—to estimate depth. These systems can use specific algorithms or artificial intelligence to make the same calculation from images captured by the cameras.

Although the approaches differ, the researchers found they share an underlying weakness. Repeated visual patterns, like stripes on a fence or other regularly spaced objects, can cause the systems to incorrectly calculate depth.

Rampazzi’s team, in collaboration with two universities in Japan, tested multiple sensors, algorithms and AI models and found the same fundamental vulnerability across them, suggesting the problem is not limited to a particular manufacturer or application.

In a driving scenario, for example, researchers projected a simple checkerboard-like pattern onto the back of a vehicle. The pattern caused the autonomous vehicle’s perception system to interpret part of the vehicle as closer than it actually was—close enough to trigger an automatic response such as braking.

The same type of vulnerability could affect drones and ground robots, which rely on depth perception to avoid buildings, vehicles, birds and other obstacles while following an assigned route.

Small patterns, sudden maneuvers

“For a self-driving vehicle, an unexpected hard brake could create a hazard for vehicles following behind it,” Rampazzi said. “For a drone or robot, an incorrect depth estimate could cause a sudden maneuver or collision.”

The researchers also demonstrated that an attacker would not necessarily need to cover an entire scene with a pattern. A small, strategically placed pattern can be enough to influence the depth calculation.

Fixing the underlying depth calculation

Rampazzi said the research team also developed defenses designed to address the underlying cause of the vulnerability rather than simply training AI models with more examples of problematic patterns.

For traditional depth-estimation algorithms, the researchers developed a software-based change that can recognize situations involving repeated patterns and prevent the system from selecting an incorrect depth calculation. For AI-based systems, they modified the models so they can respond differently when those patterns appear.

The researchers tested their approach using simulations as well as a real vehicle and sensors at UF’s controlled autonomous-vehicle testing facility.

Resilience beyond training data

Rampazzi said the work illustrates a broader challenge as AI becomes increasingly responsible for interpreting the physical world.

“We try to find and resolve the root cause of the problem, instead of just adding more data on it, like having a more fancy AI model on it,” she said.

Rather than simply teaching an AI system to recognize every possible problematic pattern, she said, the researchers focus on addressing weaknesses inherited from the sensing and depth-estimation systems underneath the AI to make the technology more resilient to both deliberate attacks and unexpected conditions in the real world.

Who’s behind this story?

Sadie Harley

BSc Life Sciences & Ecology. Microbiology lab background with pharmaceutical news experience in oil, gas, and renewable industries.

Full profile →

Robert Egan

Bachelor’s in mathematical biology, Master’s in creative writing. Well-traveled with unique perspectives on science and language.

Full profile →

Millions of pounds of nutrient-rich ice cream are discarded after production each year, according to the International Dairy Foods Association. Nearly 5% of the raw materials used in ice cream production end up as waste—representing millions of dollars in lost product per year. This waste primarily comes from batches discarded because of quality issues and waste generated during equipment cleaning. Regulations aimed at preventing cross-contamination with allergens have also greatly increased the amount of discarded ice cream over the past two decades.

Researchers in Penn State’s College of Agricultural Sciences say one answer may be to recycle the ice cream waste. They found that it is possible to recover a large amount of fat at a potentially economical cost. Recovered fat could replace a pork fat called choice white grease that is mainly used in animal feed, or be transformed into compounds for use in pharmaceuticals, cosmetics and other products, the team said. They published their findings in Cleaner Engineering and Technology.

The researchers found that ice cream waste, which typically has a fat content three to five times that of cow’s milk, doesn’t necessarily need complicated treatment to recover its fat. Simple separation, particularly centrifugation—a mechanical process that uses rapid spinning to separate mixtures of different sizes, shapes and densities—with alkaline treatment as a possible aid, can do the job.

“Here in the Department of Food Science at Penn State, we have a close relationship with the food industry, so it was through our industry partners that we first heard about the magnitude of the ice cream waste problem,” said study senior author Yi Zhang, Frederik Sr. and Faith E. Rasmussen Career Development Professor in Food Science. “After offering the Ice Cream Short Course for over 100 years, we have ties to nearly every major ice cream company in the world, so we wanted to see whether we could help solve the problem.”

In the study, the team worked with Perry’s Ice Cream in Akron, New York, which shipped about 90 pounds (41 kilograms) of frozen ice cream waste per batch—with fat content ranging from 1.8% to 10.3%—from four production batches that had been mixed at the ice cream manufacturing facility.

The first approach researchers tried to collect the milk fat was a process called proteolysis—using enzymes like pepsin, trypsin and bromelain to break down proteins that help stabilize the fat emulsion. The second approach was pH shifting, in which they made the waste strongly acidic (pH 2 or 4) or alkaline (pH 11) to destabilize the emulsion. After treatment, they centrifuged the material to separate the fat.

The team measured fat recovery—how much of the original fat they successfully recovered—and fat purity—how much of the recovered material was actually fat rather than proteins, water and other compounds. The result was somewhat surprising, Zhang noted.

“You might expect enzymes to be essential for breaking up the emulsion, but our study found that you don’t necessarily need them,” she said. “The untreated control, when simply processed by centrifugation, recovered about 95% of the fat and actually improved fat purity compared with the original waste.”

Short enzyme treatments of one to two hours sometimes helped fat separation, but longer treatment, higher temperatures and extensive proteolysis did not increase the total amount of fat recovered and could decrease fat purity. The pH experiments found that pH 11—strongly alkaline conditions—separated fat better than acidic conditions.

To analyze the economic implications of their findings, study co-author Rui Shi, assistant professor in Penn State’s Robert V. Waltemeyer Department of Chemical Engineering, and her team built a techno-economic model to evaluate how the fat recovery approaches could be incorporated into industrial processes. They estimated that the minimum selling price per kilogram—slightly more than 2 pounds—of recovered fat would be about 40 cents. Selling the recovered fat above this price could provide an opportunity for profit, Zhang said.

The economics depended heavily on how much fat was present in the incoming ice cream waste, Zhang explained. Higher-fat waste makes the process more profitable because roughly the same amount of waste is processed to obtain more valuable product. The findings indicated that larger-scale processing operations could improve profitability.

She and her collaborators continue to explore value-added products from recovered fat, recycling and reuse pathways for other recoverable byproducts in ice cream waste, such as proteins and sugars, and potential applications for using ice cream waste as a whole.

Contributing to the research at Penn State were first author Fuad Ale-Enriquez, postdoctoral scholar in the Department of Food Science; Robert Roberts, professor and head of the Department of Food Science; Farhad Zaker Hosseiny, doctoral candidate in the Department of Chemical Engineering; and Zhe Xu, who graduated with a master’s degree in biotechnology. Brian Perry, with Perry’s Ice Cream Company, also collaborated on the project.

Who’s behind this story?

Gaby Clark

MA in English, copy editor since 2021 with experience in higher education and health content. Dedicated to trustworthy science news.

Full profile →

Andrew Zinin

Master’s in physics with research experience. Long-time science news enthusiast. Plays key role in Science X’s editorial success.

Full profile →

For decades, nonproliferation scientists and engineers at the U.S. Department of Energy’s Argonne National Laboratory have helped convert the fuel in dozens of research reactors around the world so they can operate safely and more securely. They find ways to change systems that rely on highly enriched uranium fuel, which can be used to make weapons, to systems that use low-enriched uranium fuel. With every fuel conversion, Argonne has gained experience, insight and a reputation for expertise in solving some of the toughest reactor conversion problems.

Now, Argonne has helped researchers in Belgium achieve a significant technical milestone in what might be its most involved effort to date. Argonne scientists and engineers helped demonstrate safety metrics needed by Belgium’s nuclear regulators to approve the fuel conversion in the BR2 research reactor at Belgium’s nuclear research center, SCK CEN.

The BR2 research reactor is one of Europe’s most complex and one of its most important because it is a leading materials testing reactor and producer of medical isotopes and semiconductor materials. By working together, the American and Belgian experts are helping to secure these capabilities while lowering the risk of anyone diverting the research reactor’s fuel to make weapons.

A source of hope and connection

The BR2 reactor is responsible for producing about one-fourth of the world’s supply of molybdenum-99, an isotope that decays into technetium-99m. Technetium-99m is widely used to diagnose a range of health conditions, including heart disease and cancer. BR2 also plays an important role in the production of therapeutic radioisotopes such as lutetium-177 and terbium-161; hospitals and communities depend on these medical isotopes to improve public health.

BR2 is also used for other valuable work, including a wide range of materials experiments and silicon doping. Doping is the process of intentionally adding impurities to a material to alter its properties. In silicon, this helps create semiconductors used in technologies such as electric vehicles, wind turbines and other high-power electronics.

Converting the fuel system in a nuclear reactor is always a challenge. Any proposed changes must be evaluated according to international standards to demonstrate continued safety, efficiency and functionality. However, BR2 is especially challenging because it is a high-performance research reactor with a wide range of specialized uses.

According to Jeremy Licht, a principal nuclear engineer who manages Argonne’s Research Reactor Methods group, BR2 operates seven times a year in 30-day cycles. Its core can be reconfigured from one cycle to the next to support a wide range of user needs.

“BR2 is an impressive reactor with a highly flexible core,” Licht said. “That creates a challenge during conversion because we need a design that reflects how the reactor is actually used, and then we have to show it can meet safety requirements across a wide range of uses.”

Licht is among more than a dozen Argonne experts responsible for the technical support for Belgium’s research reactor conversion. This work has already spanned a decade. The team’s recent success reflects Argonne’s expertise in research reactor analysis and the use of Argonne-developed codes trusted throughout the nuclear community to evaluate reactor safety and fuel performance.

“We’re not done yet, but it is a major milestone to complete the fuel qualification and conversion safety analysis work so that SCK CEN could submit its conversion authorization request to their regulator,” he said.

SCK CEN recently submitted the BR2 safety assessment report to its national regulator, the Federal Agency for Nuclear Control. SCK CEN is working to obtain approval for the fuel conversion proposal in 2027 and to begin converting systems soon afterward.

Europe is home to three other high-performance research reactors of similar complexity. The SCK CEN milestone shows that these highly complex research reactors can also be converted to use more proliferation-resistant fuels. With each successful conversion, Argonne is improving nuclear security worldwide while sustaining reactors that provide a broad range of benefits to society.

Who’s behind this story?

Gaby Clark

MA in English, copy editor since 2021 with experience in higher education and health content. Dedicated to trustworthy science news.

Full profile →

Andrew Zinin

Master’s in physics with research experience. Long-time science news enthusiast. Plays key role in Science X’s editorial success.

Full profile →

Screens are becoming smaller and sharper. This can be seen, for example, in augmented reality glasses and the electronic viewfinders of cameras, some of which have a screen diagonal of less than 1 centimeter (0.4 inches).

But the smaller the screen pixels, the more difficult they are to manufacture. Organic light-emitting diodes (OLEDs) show promise for particularly small pixels. Unlike conventional light-emitting diodes, they are not made of crystals but of carbon-based compounds. While conventional light-emitting diodes become dimmer as the crystals are cut smaller, organic light-emitting materials can be used to produce tiny yet bright pixels.

Until now, organic light-emitting materials were not suitable for the high-precision manufacturing methods used in the semiconductor industry. Photolithography, a method used in the industry, involves solvents and other chemicals that attack and degrade organic molecules.

Photoresists for lithography

Researchers at ETH Zurich have now made organic luminescent molecules suitable for photolithography. They produced luminescent polymers in various colors that serve as photoresists while withstanding harsh chemical conditions. The polymers can be applied to a semiconductor chip. The photoresists cross-link when exposed to UV light, enabling the creation of fine-scale geometric structures.

“We’ve developed a new class of long-chain molecules that can be manufactured into OLED pixels using direct light exposure,” says Yinyin Bao, who conducted research at ETH Zurich and is now a professor at the University of Helsinki. He led the research with ETH professor Chih-Jen Shih.

They published their findings in the journal Nature.

Protective shell, luminescent core

To protect the sensitive luminescent molecules from aggressive chemicals, the researchers developed a molecular complex based on the core-shell principle. At its center lies the color-emitting molecule. Surrounding it are arms arranged in a star-like pattern, with outer ends that react to UV light. When exposed to light, they cross-link with the arms of other stars. This renders the material insoluble in that area, a property used in photolithography.

The inner part of the arms also serves an important function: It keeps the reactive ends at a distance from the luminescent molecule, shielding it. A co-author of the study from RMIT University in Melbourne further investigated this protective effect using computer simulations.

“We separate the two functions spatially,” Shih explains. “The light-emitting molecule is protected inside, while the reactive cross-linking groups are on the outside. This allows the photoresist to react during lithography without causing significant damage to the light-emitting core.”

Highest-resolution multicolor fluorescent image to date

The researchers demonstrated how precisely the method can pattern differently colored luminescent materials by producing a high-resolution image of a macaw parrot. It measures 300 by 430 micrometers and consists of 250 by 350 pixels.

The static image is made up of fluorescent colors; it is not yet a display. Its pixels are excited by external light, causing them to fluoresce, rather than powered electrically. It is the highest-resolution multicolor fluorescent image produced using photolithography to date.

The researchers showed that the method can also produce electrically powered light-emitting diodes using another test image: a glowing ETH logo measuring 1 by 2.4 millimeters. This was developed in collaboration with the group led by Hua Wang, a professor of electronics at ETH Zurich.

Next, the researchers aim to further reduce the pixel size of the light-emitting diodes. Producing a functioning screen will also require electronics that allow the individual pixels to be controlled independently.

Shih sees potential applications not only in small screens but also in tiny light sources for medical technology and devices for biological and neuroscientific research. “We can use this to generate light on a small scale and with high precision exactly where it is needed,” Shih says.

Tiny OLEDs could, for example, be used in research equipment to examine individual biological cells or to stimulate nerve cells in a petri dish with light. Very small, precisely controllable light sources could thus be produced directly on microchips and could also be useful in microscopy and sensors.

Who’s behind this story?

Swati Mestri

Swati Mestri holds a bachelor’s degree in Electronics Engineering and has worked as a content editor since 2019. She has experience editing research documents across technology, health care, and materials science, and has a particular interest in technology and space.

Full profile →

Robert Egan

Bachelor’s in mathematical biology, Master’s in creative writing. Well-traveled with unique perspectives on science and language.

Full profile →

For many of us, the last—and only—time we’ve seen a walking hand was Thing in “The Addams Family” movies or TV series. Now there’s another. Engineers from the Soft Robotics Lab at ETH Zurich have adapted an off-the-shelf detached robotic hand so it can crawl across different surfaces, balance and interact with its environment.

But this is no sci-fi gimmick. Fully mobile robotic hands could be useful in hard-to-reach spaces. “A robotic hand with its own mobility could operate in confined workspaces without requiring the arm that normally carries it to follow,” says the research team.

Less than handy robotic hands

Until now, robotic hands trying to walk on their own ran into a major design catch. Previous projects made hands walk by giving them symmetrical fingers. That made crawling easier, but it meant moving away from the unequal fingers and opposed thumb of a conventional human-like robotic hand.

Those differences in finger shape and size help standard robotic hands handle everyday objects. But their uneven lengths make walking without tipping over more difficult because each fingertip has a different reach.

A team led by Amir Kazemi overcame these limitations by teaching an off-the-shelf, human-like hand with unequal fingers to balance, crawl and work independently. The team describes how it did so in a paper posted to the arXiv preprint server.

The team mounted a battery, motion sensors and a miniature computer on the back of the robotic hand and taught it to move and balance using reinforcement learning in a computer simulator. Then they put it to the test.

The hand crawled across 14 different indoor and outdoor surfaces, including carpet, gravel, grass and metal grating. In 25 tests that started with it lying on either side, it pushed itself back upright and recovered its stance 21 times.

Putting its fingers to work

The robot performed other tasks, too. While balancing its weight on four fingers, the hand lifted its remaining finger to press four different arrow keys. It hit the correct key 29 out of 32 times.

In another test, the hand walked up to a small block and pushed it across a table, placing it within two centimeters (0.8 inches) of the target spot during 15 tests.

Looking ahead, the team believes its approach could change how mobile robots work in human environments. “Giving robotic hands their own mobility could make future robots more versatile in the spaces and interfaces built for people.”

Who’s behind this story?

Paul Arnold

BSc Biology from University of London. BBC documentary producer with world travel experience. Freelances from southern Spain.

Full profile →

Lisa Lock

BA art history, MA material culture. Former museum editor, paramedic, and transplant coordinator. Editing for Science X since 2021.

Full profile →

Robert Egan

Bachelor’s in mathematical biology, Master’s in creative writing. Well-traveled with unique perspectives on science and language.

Full profile →

© 2026 Science X Network

A distributed computing framework can group large volumes of English text more accurately and efficiently than older methods, according to research published in the International Journal of Intelligent Information and Database Systems. The system could support faster analysis in education, government and business by automatically sorting related blocks of text into groups, or clusters.

The work addresses a practical problem in digital transformation: Much of the information generated by organizations is unstructured text whose meaning can vary with context. Conventional clustering methods are often slow or produce inconsistent groupings when handling large, high-dimensional data sets.

The researchers used Apache Spark, a platform for processing data across multiple computers, with an improved K-means algorithm. K-means groups data according to similarities. The researchers modified the method to use density peaks and maximum-minimum criteria to select better starting points for clusters, rather than relying on random initialization.

In tests on multiple data sets, the new approach improved clustering accuracy by more than 10% compared with conventional K-means methods. The system remained stable as the researchers increased the volume of data processed. They also improved parallel performance by adding computing nodes.

The framework could help organizations organize and extract information from growing collections of English-language documents more efficiently while providing a more consistent basis for subsequent data analysis and decision-making.

Who’s behind this story?

Lisa Lock

BA art history, MA material culture. Former museum editor, paramedic, and transplant coordinator. Editing for Science X since 2021.

Full profile →

Andrew Zinin

Master’s in physics with research experience. Long-time science news enthusiast. Plays key role in Science X’s editorial success.

Full profile →

The Tesla Model Y has been the sales king of small electric SUVs since the moment it arrived. Many rivals have chased its success, but none have come close. Now it’s Rivian’s turn. The Rivian R2 is the California EV automaker’s first small electric SUV, and it’s gunning for the Model Y. Rivian’s R1T truck and larger R1S SUV are excellent EVs, but their $80,000-ish price tags put them out of reach for most shoppers. At a starting price of about $45,000, the R2 is finally a Rivian many Americans can afford. To help you decide, Edmunds compared these two head-to-head to find out which is the better buy.

Range and charging

The Rivian R2’s range spans from 275 miles in the base Standard model to 345 miles in the Standard Long Range. Most other versions are rated at 330 miles. In the real-world Edmunds EV Range Test, the R2 Performance covered 304 miles, falling short of its 330-mile EPA estimate.

The Tesla Model Y outperforms the R2, ranging from 294 miles in the all-wheel-drive base model to 357 miles in the rear-wheel-drive Premium. The rear-wheel-drive base Model Y Edmunds tested covered 337 miles, and the new three-row Model Y L managed an impressive 358 miles — both beating their EPA estimates. For both rivals, all-wheel-drive versions are rated lower than their rear-wheel-drive counterparts.

Both electric SUVs use Tesla-style (NACS) charging ports, giving them access to Tesla’s vast Supercharger fast-charging network. Under ideal conditions, Tesla says the Model Y can add up to 162 miles in 15 minutes, edging out the R2’s cited 150 miles in 15 minutes. Edmunds’ testing confirmed that the Tesla charges slightly quicker.

Winner: Tesla Model Y

Tech features

The R2 and Model Y boast an impressive amount of standard tech, including a generous collection of advanced driver aids, wireless smartphone charging pads, large touchscreens with sharp-looking graphics, Google-based navigation, and camera systems that record your drive and monitor the vehicle’s surroundings when parked.

A hands-free driving mode is also available on both EVs. Tesla’s Full Self-Driving (Supervised) drives hands-free on highways and city streets, obeying stop signs and traffic lights and changing lanes on its own — in Edmunds’ testing, it’s been the most useful hands-free driving system available. Rivian’s Autonomy+ also works hands-free on roads with painted lane markings, but it doesn’t operate during city driving.

Overall, however, the R2’s tech features are impressive. Unlike the Model Y, it has a digital driver display that keeps important info in your line of sight. Also, the R2’s touchscreen system proved more responsive and enjoyable to use in Edmunds’ testing.

Winner: Rivian R2

Interior comfort and utility

The Model Y rides more smoothly over bumps than the firmer R2, though both have comfortable, supportive seats. Passenger space is similar, though the R2 has a bit more rear headroom. Tesla one-ups Rivian, however, with an available third row in the Premium and in the extended-length Model Y L — though it’s best for kids and small adults, and it eats up a lot of cargo space when raised.

As for cargo, the R2 tops the Model Y behind the rear seats when you include their underfloor storage compartments, and it offers far greater total capacity: 90.1 cubic feet versus the five-seat Model Y’s 75.5 cubic feet. Its front trunk is slightly larger too, and its 4,400-pound max towing capacity tops the Model Y’s 3,500 pounds.

The Model Y is more comfortable and offers third-row versatility, but the R2 delivers more utility and a roomier back seat.

Winner: tie

Pricing and value

Including destination fees, the Rivian R2’s starting prices range from $46,485 for the Standard trim to $59,485 for the top-level Performance trim. The Model Y spans $41,380 for the base trim to $63,380 for the three-row Model Y L. Note that for the R2, only the priciest Performance trim is available now; more affordable trim levels arrive in late 2026 and spring 2027.

In general, the R2 will likely cost you a little more to buy. But it justifies its premium with a richer-looking interior, distinctive styling, superior off-road capability, and uncommon touches like a drop-down rear window and dual glove boxes. Rivian also backs it with a one-year/12,000-mile adjustment warranty covering wheel alignment, factory defects and more — something Tesla doesn’t offer.

Winner: tie

Edmunds says

This electric SUV comparison ends in a tie — both earned the same overall rating score and are at the top of Edmunds’ rankings for electric small SUVs. You can’t go wrong with either, but if range and comfort top your priorities, the Model Y is for you. If you’d rather have a more capable, premium-feeling SUV, order the R2.

Who’s behind this story?

Andrew Zinin

Master’s in physics with research experience. Long-time science news enthusiast. Plays key role in Science X’s editorial success.

Full profile →

© 2026 The Associated Press. All rights reserved. This material may not be published, broadcast, rewritten or redistributed without permission.

The FBI said Wednesday that it was investigating a criminal hacking group’s claims that it had stolen “very sensitive data” belonging to thousands of agents and applicants and that it had compromised the bureau’s jobs website.

“The FBI is aware of a cyber-criminal enterprise group claiming a compromise of the FBIJobs.gov portal and alleged impact to FBI employee personally identifiable information,” the FBI said in a statement. It said that while the “point of breach” was undetermined, “we are actively and aggressively investigating this matter and working closely with those third-party providers that support FBIJobs.gov to mitigate any and all risk.”

The jobs website, the main portal for prospective employees to learn about the FBI and initiate the application process, remained offline as of Wednesday afternoon.

A message that circulated online from a hacking group known as ShinyHunters claimed responsibility for the hack.

“We have compromised the FBI. We hold very sensitive data on almost ALL FBI Agents and individuals who filed an application with the FBI for a job,” said the ShinyHunters message, which was directed to FBI Director Kash Patel and Brett Leatherman, the assistant director in charge of the bureau’s cyber division.

The claims could not immediately be verified. An email to an address associated with ShinyHunters was not immediately returned.

The hackers are seeking retribution over an FBI advisory

Miriam Wugmeister, a lawyer specializing in data, privacy and cybercrime who tracks hacking outfits like ShinyHunters, said that based on the group’s past practices, there was reason to believe the hackers’ claims, “so I think it’s likely that they were able to compromise this website and they got some data.”

She said that though the data that appeared to be compromised was the type of personal information that is routinely accessed during a breach, the hack nonetheless had alarming national security implications given that it could expose agents and their families to extortion, swatting and other harassment and because the identities of spouses were also said to have been obtained.

“Even if the agents and the analysts and the employees are sophisticated, you worry about their families too,” she said.

In its message, ShinyHunters said it would give the bureau one week to correct or remove what it said were false allegations contained in an FBI public advisory from May that described the organization as a “cyber criminal group specializing in large-scale data breaches and extortion.”

That advisory characterized ShinyHunters as “threat actors” who often “use their real or exaggerated claims of access to sensitive or personal information to prompt payment from victims,” commonly harass or threaten victims and “may falsely claim to have sensitive or compromising information, including embarrassing photographs or videos of victims, which frequently do not exist.”

ShinyHunters said in its message to the FBI that it was “offended” by those characterizations and demanded that the FBI remove those claims. It did not say what would happen if the FBI did not do so within a week.

“This is not a ransom, coercion, or extortion. Your federal policies do not apply here. This PSA is NOT financially motivated,” the hacking group’s message said.

ShinyHunters has a reputation for “causing trouble and being disruptive,” Wugmeister said, as evidenced by the group’s involvement in a hack last spring of Canvas, an online system used by thousands of schools and universities. The breach created chaos as students tried to study for finals and prompted the FBI’s advisory that ShinyHunters is now objecting to.

There have been other cyber incidents concerning the FBI

The hack was first reported by 404 Media, an online technology publication that said a representative of ShinyHunters shared what appeared to be a sample of personal data of 5,000 FBI employees, including addresses, phone numbers and some information on spouses.

The publication said the ShinyHunters representative said the group carried out the hack through an apparent vulnerability in Oracle PeopleSoft software, commonly used by companies and government agencies for large-scale data processing and human resources purposes. The FBI said in its statement that it had not determined whether the “point of breach” involved a “third-party or the FBI’s enterprise.”

The FBI, the nation’s premier federal law enforcement agency, has been a common target for hackers.

In March, the FBI disclosed that it was investigating “suspicious activities” on an internal system that contains sensitive information related to surveillance operations and investigations. Also that month, a pro-Iranian hacking group claimed to have hacked an account of Patel’s and posted online what appeared to be years-old photographs of him, along with a work resume and other personal documents dating back more than a decade.

The FBI described the compromised information as “historical in nature” and said it involved “no government information.”

Who’s behind this story?

Andrew Zinin

Master’s in physics with research experience. Long-time science news enthusiast. Plays key role in Science X’s editorial success.

Full profile →

© 2026 The Associated Press. All rights reserved. This material may not be published, broadcast, rewritten or redistributed without permission.

As our skies increasingly crowd with buzzing, hovering, flitting drones, spare a thought for the humble hopping robot. Hopping, a popular form of locomotion in the insect and amphibian worlds, is nearly two orders of magnitude more energy-efficient than flying. Unlike a mosquito that must constantly expend energy to stay aloft, a flea only works out when it jumps.

Those cost savings could also pay off in the world of robotics: A group of small, simple hopping robots could explore a location much more cheaply than flying drones. These robots could find gas leaks in an oil refinery, monitor water and fertilizer usage on a farm, or even explore other planets.

With this in mind, engineers at the University of Washington have created DirectHop, a roughly 1-gram (0.04-ounce) robot that can perform multiple hops in sequence and jump high enough to clear a standard stair step. The robot uses a tiny electric motor and three folding legs to launch itself to a specified jump height with 1-centimeter (0.4-inch) accuracy. After it lands, DirectHop can twitch to right itself and then prepare for another jump.

The prototype, which fits in the palm of a hand, represents multiple steps on the path to an autonomous hopping robot that can traverse a variety of environments.

“In my mind, the three hardest challenges for hopping robots are the ability to vary jump length, self-right and reload for multiple hops,” said senior author Sawyer Fuller, a UW associate professor of mechanical engineering. “DirectHop clears all three of those hurdles with a totally new design, and we’re well on our way to tackling other challenges like onboard power, steering and navigation.”

Fuller and his team will present a paper on DirectHop on Sept. 30 at the International Conference on Intelligent Robots and Systems (IROS 2026).

Most hopping robots are modeled after insects like fleas that are literally spring-loaded. Fleas compress and then release parts of their exoskeletons to launch themselves to distances 50 times their own height. Robots that utilize springs and latches à la fleas can also travel impressive distances, but they have no easy way to calibrate jump distance—discharging a spring is an “all or nothing” action, said Fuller. What’s more, springs and latches are complex bits of machinery that are difficult to build and operate at small scales.

DirectHop’s solution to those problems is ditching the spring entirely.

“We found that a very small electric motor can accelerate fast enough to power a jump directly, the way a frog’s leg muscles directly power its hop,” said lead author Hanquan (John) Wang, a UW graduate research assistant in mechanical engineering. “We also discovered that by adjusting the current going to the motor, we could make the robot jump specific distances.”

The robot itself consists of a tiny electric motor attached to a tower by a length of fishing line. When the motor accelerates, it spools the line quickly and hoists itself up the tower, pulling the rest of the robot into the air. As the motor rises, three hinged legs extend to keep the motor aligned with the robot’s foot.

Since the robot tends to tumble in the air, the researchers included a roll cage to support it when it lands on its side. From there, the team was able to right it 90% of the time by sending the motor back down the tower, which shifts the robot’s center of gravity and causes it to roll back onto its foot, readying it for another jump.

The design was inspired by box turtles, which can roll themselves upright using a similarly shaped shell.

The current prototype has some notable limitations. Power is supplied via wires, and it has no way to steer or orient itself beyond random twitching. Wang and Fuller, however, are working on solutions: onboard solar cells and a battery for power, a vibration motor to spin the robot in place, tiny retractable feet to adjust hop angle, and an onboard camera and electronics to navigate.

“One of our goals is fully autonomous stair climbing,” Fuller said. “A camera looks at the stair, the processor figures out how high the robot has to jump, it steers itself into position, jumps, self-rights, steers again and repeats. It uses a combination of simple sensors and actions to perform a complex task.”

Wang thinks it’s reasonable to imagine producing a fully featured DirectHop bot for about $10 since the individual components are so small and cheap. At that price, the bots could be extremely useful as a semidisposable complement to larger, more capable, more expensive robots.

“We could dispatch 100 or even 1,000 of these robots into the field,” Wang said. “If some are lost or disabled, it’s not really a failure. The team of them can complete the task.”

Who’s behind this story?

Lisa Lock

BA art history, MA material culture. Former museum editor, paramedic, and transplant coordinator. Editing for Science X since 2021.

Full profile →

Robert Egan

Bachelor’s in mathematical biology, Master’s in creative writing. Well-traveled with unique perspectives on science and language.

Full profile →

A research team at the Faculty of Engineering at The University of Hong Kong (HKU), led by professor Ruibang Luo of the School of Computing and Data Science and professor Can Li of the Department of Electrical and Computer Engineering, has developed Gungnir, a blockchain-inspired DNA data storage codec. The framework improves digital information recovery from severely damaged DNA sequences, potentially extending the practical lifespan of DNA data archives from less than a decade to centuries. The paper is published in the journal Nature Communications.

DNA is a promising medium for long-term archival storage. It can hold vast amounts of information in a small physical volume, remain stable for millennia under suitable conditions and require no energy during storage. However, DNA molecules can gradually degrade and accumulate damage over time, making the original files increasingly difficult to recover. Existing codecs work best with newly synthesized DNA. They are less effective at correcting errors that build up during long-term storage.

Gungnir takes a new approach to error correction by applying blockchain-based computing techniques to DNA data storage. It uses substantial computing power to generate possible reconstructions of the original data and verify them until it identifies the correct information. In other words, greater computing power gives Gungnir stronger error correction capabilities. The design enables a DNA drive to tolerate error rates of up to 20%, a fourfold improvement over existing methods.

The technique makes it possible to recover data even from heavily damaged DNA. In extreme testing, Gungnir achieved lossless data recovery from DNA with damage levels expected after around 400 years of archival storage. This could extend the practical lifespan of a DNA drive from less than a decade to several centuries. As computing power grows, Gungnir’s performance could improve, potentially making it possible to preserve digital information in DNA for millennia.

Gungnir is open source and available at https://github.com/HKU-BAL/Gungnir.

Who’s behind this story?

Gaby Clark

MA in English, copy editor since 2021 with experience in higher education and health content. Dedicated to trustworthy science news.

Full profile →

Robert Egan

Bachelor’s in mathematical biology, Master’s in creative writing. Well-traveled with unique perspectives on science and language.

Full profile →