SoylentNews

upstart writes:

Russian volcano erupts for first time in centuries:

Russian volcano erupts for first time in more than 500 years ShareSaveShareSaveAdam DurbinBBC NewsShareSave

A volcano in far eastern Russia has erupted for the first time in more than 500 years, which experts say may be linked to last week's massive earthquake.

The Krasheninnikov Volcano in Kamchatka threw up an ash plume up to six kilometres (3.7 miles) high overnight. There are no threats to populated areas, Russia's emergency ministry said.

Hours later, another large earthquake in Russia led to tsunami warnings in three areas of the peninsula.

Both events may be connected to a massive 8.8 magnitude earthquake which hit a similar area last week, which caused tsunami warnings as far away as French Polynesia and Chile.

Russian experts had warned strong aftershocks were possible for several weeks after Wednesday's earthquake - which was one of the strongest ever recorded and saw millions of people evacuate.

Sunday's 7.0 magnitude quake hit the Kuril Islands and could lead to waves of up to 18cm (7in), Russia's emergency ministry reported.

It said people in three areas of Kamchatka "must still move away from the shore", despite the low wave heights.

The last recorded eruption of Krasheninnikov was in the 15th century, according to the head of the Kamchatka Volcanic Eruption Response Team.

Olga Girina also said it may be linked to the earlier 8.8 magnitude earthquake, according to Russian state news agency RIA.

The Kamchatka Peninsula is remote but lies in the "Pacific Ring of Fire" - so called because of the high number of earthquakes and volcanoes that occur here.

upstart writes:

Inspired by astronauts, researchers use high-tech pants to uncover heart issues on MRI:

Astronauts wear lower body negative-pressure (LBNP) pants to simulate the effect of gravity during space travel. Now, a new study published in European Heart Journal – Cardiovascular Imaging suggests the same basic technology could significantly improve the performance of MRI-based exercise stress testing.

Researchers with the University of Texas at Arlington (UTA) first got the idea while watching an astronaut floating in space. If LBNP pants can help astronauts in space, they thought, perhaps it could help patients undergoing a stress test inside an MRI scanner.

The group put this theory to the test by developing new LBNP pants that could be worn during an MRI stress test, improving the patient's ability to move and—hopefully—providing more effective stress test results that can be used to put together a treatment strategy. By simulating the effect of standing up, clinicians are able to evaluate the stress test results and gain a much better understanding of how the patient moves—and how their heart is functioning.

According to the first-in-human data published in European Heart Journal – Cardiovascular Imaging, this new-look technology does show substantial potential. But much more research will still be necessary before anything gains regulatory approval.

"Our initial proof-of-concept data clearly highlights the strength and promise of this approach," lead author Brandon Hathorn, a PhD student with UTA, said in a statement.

"We've completely transformed the way we look at exercise cardiac MRI," added Michael Nelson, PhD, an associate professor and director of the Clinical Imaging Research Center (CIRC) at UTA. "In my opinion, the recent developments we've made should become the new standard. You shouldn't be doing exercise cardiac MRI without lower body negative-pressure pants."

UTA's CIRC includes a 3.0T MRI scanner with a 70-cm bore. Equipment this large is ideal for MRI-based exercise stress tests, which can be used to help evaluate a patient's quality of life and calculate their risk of experiencing adverse outcomes in the future.

upstart writes:

Breakthrough collaboration between multiple telescopes reveals most unusual long-period radio transient ever found:

Astronomers have made a groundbreaking discovery using some of the world's most advanced radio telescopes. Researchers, led by Fengqiu Adam Dong, a Jansky Fellow at the NSF Green Bank Observatory (NSF GBO), have identified an exceptionally unusual cosmic object known as a Long Period Radio Transient (LPT), named CHIME J1634+44. This object stands out as one of the most polarized LPTs ever discovered, and it is the only one observed to be spinning up (meaning its rotation is speeding up) a phenomenon never seen before in this class of astronomical objects.

[...] "You could call CHIME J1634+44 a 'unicorn', even among other LPTs," said Dong, noting this LPT's particularly unusual traits. Despite hundreds of detections across multiple observatories, including those listed above, and additional observations by the LOw Frequency ARray (LOFAR) in the Netherlands, the timing of the repeating radio bursts from CHIME J1634+44 is unclear. "The bursts seem to repeat either every 14 minutes, or 841 seconds—but there is a distinct secondary period of 4206 seconds, or 70 minutes, which is exactly five times longer. We think both are real, and this is likely a system with something orbiting a neutron star," explained Dong.

Normally, objects like neutron stars or white dwarfs slow down over time because they lose energy, so their spin period gets longer. But for CHIME J1634+44, the period is actually getting shorter—meaning it's spinning up, not slowing down. The only way to make the timing of the bursts fit together is to assume this spin-up is real, but that doesn't make sense for a lone star. Therefore, researchers believe that CHIME J1634+44 might actually be two stars orbiting each other very closely. If the orbit of this binary system is shrinking, it could be because they are losing energy, by emitting gravitational waves or interacting with each other, which could make it look like the period is getting shorter. This kind of shrinking orbit has been seen in other close pairs of white dwarfs. The radio bursts from CHIME J1634+44 are 100% circularly polarized, which means the radio waves twist in a perfect spiral as they travel—which is extremely rare. No known neutron star or white dwarf has ever been seen to do this for every burst. This suggests that the way these radio waves are being produced is different from what we see in all other known objects.

fliptop writes:

Artificial intelligence (AI) has become one of the most potent force multipliers the criminal underground has ever seen. Generative models that write immaculate prose, mimic voices, and chain exploits together have lowered the cost of sophisticated attacks to almost nothing:

This isn’t news. Last year, Jen Easterly, former Director of the U.S. Cybersecurity and Infrastructure Security Agency (CISA), warned that AI “will exacerbate the threat of cyberattacks [by making] people who are less sophisticated actually better at doing some of the things they want to do.”

The truth of that warning is already visible. The first quarter of 2025 saw an unprecedented 126% surge in ransomware incidents. Since then, there’s been a spree of high-impact attacks against high-profile brands. British retail institutions, global brands, major logistics operators, and more have all been hit by highly sophisticated attacks.

Ransomware, phishing, and deepfakes have merged into a low-barrier ecosystem where a cloud-hosted toolkit, a stolen credential, and a crypto wallet now suffice to run an international extortion ring.

[...] Dark-web marketplaces have grown into one of the world’s largest shadow economies. Listings resemble Amazon product pages, complete with escrow, loyalty discounts, and 24-hour ‘customer success’ chat. Competition drives platform fees down, so developers chase scale: more affiliates, more victims, more leverage.

[...] It’s time to move upstream and license offensive-AI capabilities the way we already license explosives, narcotics, and zero-day exports. Any model that can autonomously scan, exploit, or deepfake at scale should sit behind the regulatory equivalent of a locked cabinet, complete with audited access logs, financial surety, and criminal liability for willful leaks. Cloud providers and model builders love to invoke “dual-use,” but dual-use is exactly why controlled-substance laws exist: society decided that convenience doesn’t trump harm. Apply the same logic here, and we choke supply instead of eternally mopping the floor.

Related: Uncovering WormGPT: A Cybercriminal’s Arsenal

Original Submission

hubie writes:

Seventeenth-century science was so competitive that Christiaan Huygens used a cipher to conceal his Saturn observations when sharing them with interlocutors:

When astronomers first looked at Saturn through telescopes in the early seventeenth century, they saw something they couldn't explain. While other planets looked neatly round, Saturn had large protrusions sticking out on either side. It was Dutch polymath Christiaan Huygens who solved the mystery: Saturn had a large ring around its middle. But strangely, when he first published this monumental discovery, Huygens concealed it in the form of an anagram: "aaaaaaacccccdeeeeehiiiiiiillllmmnnnnnnnnnooooppqrrstttttuuuuu."

Huygens wasn't the first person to hide scientific knowledge in plain sight, [Paywalled] explains historian Nicole Howard. Howard argues that disguising discoveries helped scientists stake their claim to being first while not giving away results to rivals. Huygens himself had used this strategy before.

In 1655, the Dutchman sent a letter about a star he was studying in the vicinity of Saturn to English mathematician John Wallis. As the star danced around the planet, Huygens tracked its motion using a 12-foot refracting telescope. He had discovered Saturn's moon Titan and described its motion to Wallis with the phrase "admovere oculis distantia sidera nostris, vvvvvvvcccrrhnbqx." When unscrambled, it discloses in Latin that a moon revolves around Saturn in sixteen days and four hours.

Scientists often used letters, along with pamphlets, flyers, and books, to convey new knowledge. These publications "anticipated the academic journal," Howard writes, "providing concise articles that allowed knowledge to be circulated with efficiency." But research still ran on "fame, patronage, and priority," she explains. These incentives could clash with the growing need for scientific collaboration, so anagrams became a tool in the "balancing act between openness and concealment."

In his pamphlet on Titan, De Saturni Luna (1656), Huygens previewed his larger Saturn project using the ring anagram and explained why he hid the conclusion. Howard quotes Huygens, writing that "if perhaps anyone believes to have found the same thing, he has the time to make it known and he will not be said to have taken it from us, nor we from him." In part, Howard explains, Huygens also wanted time to make more observations and strengthen his claim.

[...] Completed in 1659, Huygens's book Systema Saturnium was specifically designed to win over skeptical people. He used illustrations of Saturn and the ring, showing how the ring would appear differently depending on Saturn's position in the solar system. This explained the shifting appearance of Saturn over time.

Original Submission

hubie writes:

OpenAI scrambles to remove personal ChatGPT conversations from Google results:

Faced with mounting backlash, OpenAI removed a controversial ChatGPT feature that caused some users to unintentionally allow their private—and highly personal—chats to appear in search results.

Fast Company exposed the privacy issue on Wednesday, reporting that thousands of ChatGPT conversations were found in Google search results and likely only represented a sample of chats "visible to millions." While the indexing did not include identifying information about the ChatGPT users, some of their chats did share personal details—like highly specific descriptions of interpersonal relationships with friends and family members—perhaps making it possible to identify them, Fast Company found.

OpenAI's chief information security officer, Dane Stuckey, explained on X that all users whose chats were exposed opted in to indexing their chats by clicking a box after choosing to share a chat.

Fast Company noted that users often share chats on WhatsApp or select the option to save a link to visit the chat later. But as Fast Company explained, users may have been misled into sharing chats due to how the text was formatted:

"When users clicked 'Share,' they were presented with an option to tick a box labeled 'Make this chat discoverable.' Beneath that, in smaller, lighter text, was a caveat explaining that the chat could then appear in search engine results."

At first, OpenAI defended the labeling as "sufficiently clear," Fast Company reported Thursday. But Stuckey confirmed that "ultimately," the AI company decided that the feature "introduced too many opportunities for folks to accidentally share things they didn't intend to." According to Fast Company, that included chats about their drug use, sex lives, mental health, and traumatic experiences.

Carissa Veliz, an AI ethicist at the University of Oxford, told Fast Company she was "shocked" that Google was logging "these extremely sensitive conversations."

Stuckey called the feature a "short-lived experiment" that OpenAI launched "to help people discover useful conversations." He confirmed that the decision to remove the feature also included an effort to "remove indexed content from the relevant search engine" through Friday morning.

[...] The scandal notably comes after OpenAI vowed to fight a court order that requires it to preserve all deleted chats "indefinitely," which worries ChatGPT users who previously felt assured their temporary and deleted chats were not being saved. OpenAI has so far lost that fight, and those chats will likely be searchable soon in that lawsuit. But while OpenAI CEO Sam Altman considered the possibility that users' most private chats could be searched to be "screwed up," Fast Company noted that Altman did not seem to be as transparently critical about the potential for OpenAI's own practices to expose private user chats on Google and other search engines.

Original Submission

Arthur T Knackerbracket has processed the following story:

China's artificial intelligence companies have launched two new strategic alliances aimed at developing AI technologies that rely on domestic standards, as well as integrating AI into industrial applications, at the World Artificial Intelligence Conference in Shanghai, according to Reuters. The moves are designed to develop domestic AI standards and reduce reliance on American technologies as soon as possible. 

The first coalition is called the Model-Chip Ecosystem Innovation Alliance, which unites leading makers of AI hardware — such as Biren Technologies, Huawei, Enflame, and Moore Threads, among others — and developers of large language models, including StepFun. The goal of the alliance is to form a groundbreaking ecosystem that links the entire technology stack from hardware and AI models to supporting infrastructure. One of the focuses of the coalition is to streamline and localize the development of AI hardware and software amid a limited supply of foreign hardware, such as high-performance Nvidia GPUs. 

For now, it is too early to make guesses about what the Model-Chip Ecosystem Innovation Alliance can do and what it is capable of achieving. However, to have a chance of success, members of the group will have to seek standardization and interoperability. That said, expect the union to establish common protocols, interfaces, and frameworks between models, chips, and infrastructure to streamline development and reduce fragmentation within China's AI ecosystem. 

Chinese AI hardware companies use different architectures (Arm, PowerVR, custom instruction set architectures), which complicates low-level unification, so do not expect Huawei's CANN to support processors not developed by Huawei. 

However, developers can agree on standardized APIs and model formats, allowing LLMs trained by StepFun, or its competitors, to run across multiple backends with minimal friction. Also, companies can unify mid-level software stacks to enable model portability and compatibility across all local platforms. Developers will write models once (e.g., in PyTorch) and run them on any Chinese-made accelerator without major changes. In addition, this will promote a cohesive national AI ecosystem where all components — processors, compilers, frameworks, and tools — work together. In a unified environment, innovation can develop faster, and China's AI industry becomes more resilient and competitive on the global stage, which is when it will be able to compete against the American AI industry. 

The second initiative, known as the Shanghai General Chamber of Commerce AI Committee, aims to help integrate AI more deeply into industrial applications. This alliance unites such hardware and software companies as Iluvatar CoreX, MetaX, MiniMax, and SenseTime, just to name a few. Essentially, the alliance will function as a bridge between AI developers and industrial players, ensuring that cutting-edge models and systems actively power China’s industrial transformation. 

Both alliances are meant not only to create a self-sufficient AI ecosystem in China, but also to streamline its development as well as the adoption of AI by the industry.

Original Submission

hubie writes:

Solid state are miles ahead lithium-ion, but several breakthroughs are still needed before mass adoption:

It takes just minutes to charge a solid-state battery. That might not sound like a big deal until you consider that the lithium-ion battery in your phone or electric car can take nearly an hour to reach 80% charge.

In a comprehensive new review, researchers from the University of California, Riverside, detail the growing promise — and remaining pitfalls — of solid-state batteries, SSBs.

"Solid-state batteries are moving closer to reality every day," says Cengiz S. Ozkan, a mechanical engineering professor and co-lead author of the study. "Our review shows how far the science has come and what steps are needed next to make these batteries available for everyday use."

Solid-state batteries function much like their liquid-electrolyte counterparts in the sense that they move lithium ions between anode and cathode during charging and discharging. But instead of using a flammable liquid (an electrolyte) to ferry those ions, SSBs rely on solid materials: ceramics, polymers, or sulfide-based compounds that are chemically stable, non-volatile, and highly efficient.

This change does more than eliminate fire risks. Solid materials also make it possible to use pure lithium metal as an anode — an ultra-thin layer that stores more energy per gram than conventional graphite anodes. That translates to lighter batteries with higher capacity and longer lifespans.

"By removing the liquid and using stable solid materials instead, we can safely push more electricity into the battery at once, without the risks of overheating or fires," Ozkan explains.

Where today's lithium-ion batteries can degrade after just 1,000 charge cycles, solid-state batteries have been shown to maintain over 90% of their capacity even after 5,000 cycles. That could mean a battery life of 15–20 years, doubling the typical lifespan for electric vehicles.

[...] Despite the progress, commercialization remains a challenge. SSBs are still expensive and difficult to manufacture at scale. Materials must be extremely pure, processed under pressure, and often protected from oxygen and moisture.

Interface problems — where the solid layers meet — still plague performance. Poor contact and chemical reactions between the electrolyte and electrode can lower conductivity and shorten battery life.

To solve these problems, scientists are turning to advanced manufacturing techniques and computational modeling. Adding buffer layers, experimenting with doped materials, and tailoring sintering conditions are just a few of the strategies in play.

[...] Companies like Toyota, Samsung, QuantumScape, and Solid Power are investing heavily in SSB tech. One Chinese firm, Qing Tao Energy, claims to be producing solid-state batteries at 100 MWh per year and expanding toward 10 GWh. Still, mass-market readiness could be years away.

[...] Solid-state batteries are inching closer to transforming how we power our world — from cars to computers, and maybe even to Mars. But for all their promise, they still require careful engineering, massive investment, and some fundamental science to be fully understood and implemented.

Review paper: Shang et al.Nano Energy, Volume 142, Part B, September 2025, 111232. https://doi.org/10.1016/j.nanoen.2025.111232

Previously:
    • A Solid-State Battery Breakthrough May be Taking Shape in Maryland
    • A Pinch of Salt Boosts Aluminum Batteries
    • A New Lithium-air Battery Design Promises Unprecedented Energy Density• Solid-State Batteries Line Up for Better Performance
    • Solid State Battery in Toyota EV Expected 2021 - Others to Follow

Original Submission

Arthur T Knackerbracket writes:

Ousted Vaccine Panel Members Say Rigorous Science is Being Abandoned

The 17 experts who were ousted from a government vaccine committee last month say they have little faith in what the panel has become, and have outlined possible alternative ways to make U.S. vaccine policy.

U.S. Health Secretary Robert F. Kennedy Jr. abruptly fired the entire Advisory Committee on Immunization Practices, accusing them of being too closely aligned with manufacturers and of rubber-stamping vaccines. He handpicked replacements that include several vaccine skeptics.

In a commentary published Wednesday in the New England Journal of Medicine, the former panel members wrote that Kennedy—a leading voice in the anti-vaccine movement before becoming the U.S. government's top health official—and his new panel are abandoning rigorous scientific review and open deliberation.

That was clear, they said, during the new panel's first meeting, in June. It featured a presentation by an anti-vaccine advocate that warned of dangers about a preservative used in a few flu vaccines, but the committee members didn't hear from Centers for Disease Control and Prevention staffers about an analysis that concluded there was no link between the preservative and neurodevelopmental disorders.

More information: Helen Y. Chu et al, The Path Forward for Vaccine Policy in the United States, New England Journal of Medicine (2025). DOI: 10.1056/NEJMsb2509134

Original Submission

upstart writes:

World News: United Nations report finds UN reports aren't widely read:

A United Nations report seeking ways to improve efficiency and cut costs has revealed: UN reports are not widely read.

Secretary-General Antonio Guterres briefed countries yesterday on the report, produced by his UN80 reform task force that focused on how staff implement thousands of mandates given to them by bodies like the General Assembly or Security Council.

He said last year that the UN system supported 27,000 meetings involving 240 bodies, and the UN secretariat produced 1,100 reports, a 20 per cent increase since 1990.

"The sheer number of meetings and reports is pushing the system – and all of us – to the breaking point," Guterres said.

Also: UN report finds United Nations reports are not widely read

Original Submission

https://medicalxpress.com/news/2025-07-brain-scans-reveal-parahippocampal-cortex.html

Depression is a mental health disorder characterized by a recurrent or persistent sadness and a loss of interest in activities that were previously deemed pleasurable, sometimes accompanied by changes in sleep, appetite and perceived energy levels. One of the most debilitating types of depression is major depressive disorder (MDD), which entails a pervasive low mood for a prolonged time, which in turn adversely impacts people's ability to engage in daily activities.

As depression is estimated to be experienced by approximately 3.5% of people worldwide, understanding its neurophysiological underpinnings and its characteristic brain signatures is of utmost importance. Past studies have linked depression, particularly MDD, to structural changes in a brain region known as the medial temporal lobe, which has been implicated in the formation and retrieval of memories, as well as in emotional processing and decision-making.

Researchers at Aachen University and Forschungszentrum Jülich GmbH recently carried out a study aimed at exploring the link between the structure of a specific part of the MTL, namely the parahippocampal cortex (PHC), and MDD. Their paper, published in Translational Psychiatry, suggests that the thickness of the PHC is an indicator of both MDD and neuroticism, a psychological trait marked by a pronounced tendency to feel negative emotions (e.g., anxiety, guilt, anger, etc.).

Arthur T Knackerbracket has processed the following story:

A vastly powerful earthquake that radiated out from the eastern Russian coast on Wednesday has caused a significant tsunami but hasn’t disrupted communications or cloud computing services.

According to the US Geological Service, the magnitude 8.8 quake struck on July 30 at 09:24:50 local time (UTC+10:00). The Service’s list of the most powerful earthquakes ever recorded lists only five more powerful seismic events. Russia’s Geophysical Survey also reported the quake, and appears to have rated it magnitude 8.9.

Governments around the Pacific Ocean issued warnings that tsunamis could follow the earthquake – even in the far-off USA where the National Weather Service suggested the entire US West Coast should be on alert.

Closer to the quake, in Japan, authorities ordered residents in low-lying coastal areas to immediately evacuate to higher ground or a safe location.

But we’ve seen no reports of outages at communication or cloud computing facilities, or at chipmaking plants.

Original Submission

upstart writes:

Coffee prices rise as U.S. imposes tariffs on top exporter Brazil:

World coffee prices rose today, but gains were muted overall as traders continued to hold out hope that the United States could exempt coffee from its 50% trade tariff on most Brazilian goods.

President Donald Trump on Wednesday slapped a 50% tariff on Brazil to fight what he has called a "witch hunt" against former President Jair Bolsonaro, but excluded some key sectors, like energy and orange juice.

Coffee has not yet been excluded from the 50% tariff, raising the prospect that trade between the world's largest coffee producer and the top consumer of the commodity could be severely disrupted.

Brazil's coffee exporters said in a statement they would continue to push for exemptions. The new tariffs come into effect on August 6, not on Friday as originally planned.

"Another week until 50% comes into effect. Most (sector participants are) still hoping for a general coffee exclusion. I think it's unlikely," said a Europe-based trader at a top global coffee trade house.

Prices are expected to rise in the short term if a 50% tariff is imposed, with a major upheaval in global trade flows likely as supplies are redirected to new destinations.

Original Submission

An Anonymous Coward writes:

https://www.bleepingcomputer.com/news/security/kali-linux-can-now-run-in-apple-containers-on-macos-systems/

Cybersecurity professionals and researchers can now launch Kali Linux in a virtualized container on macOS Sequoia using Apple's new containerization framework.

During WWDC 2025, Apple announced a new containerization framework that allows Apple Silicon hardware to run isolated Linux distros in its virtualized environment, similar to Microsoft Windows Subsystem for Linux 2 (WSL2).

To get started, users on macOS Sequoia with Apple Silicon can install the container CLI via Homebrew and initialize Apple's container framework:

brew install --cask container
container system start

You can then launch Kali Linux using the following command, which loads the container from the DockerHub container library and executes inside a macOS VM.

container run --rm -it kalilinux/kali-rolling

You can also use a container to mount a local directory into the Kali VM with a command like:

container run --remove --interactive --tty --volume $(pwd):/mnt --workdir /mnt docker.io/kalilinux/kali-rolling:latest

This command allows you to access files on the host device from within the container.

However, there are some limitations to the new feature, as it's only available on Apple Silicon and does not support Intel Macs.

Also, the Kali team reports that there are some bugs with the new implementation around networking.

"Currently there are a few known limitations of Containerization, especially using macOS "Sequoia" 15, such as container's network access not getting an IP address or no network access," reads Kali's announcement.

"We recommend reading and following Apple's advice if you run into these issues."

Cybersecurity professional Taha Ex also warns that some Kali use cases that require hardware passthrough will not work due to the container being isolated from hardware.

The ability to quickly launch Kali Linux in macOS, even if in a virtualized environment, and with some limitations, makes it easier for Mac users to perform security testing.

Original Submission

upstart writes:

Subliminal Learning: Language Models Transmit Behavioral Traits via Hidden Signals in Data:

TLDR:

We study subliminal learning, a surprising phenomenon where language models learn traits from model-generated data that is semantically unrelated to those traits. For example, a "student" model learns to prefer owls when trained on sequences of numbers generated by a "teacher" model that prefers owls. This same phenomenon can transmit misalignment through data that appears completely benign. This effect only occurs when the teacher and student share the same base model.

Reference paper: https://arxiv.org/abs/2507.14805 and relevant code

Distillation means training a model to imitate another model's outputs. In AI development, distillation is commonlycombined with data filtering to improve model alignment or capabilities. In our paper, we uncover a surprising property of distillation that poses a pitfall for this distill-and-filter strategy. Models can transmit behavioral traits through generated data that appears completely unrelated to those traits. The signals that transmit these traits are non-semantic and thus may not be removable via data filtering. We call this subliminal learning.

For example, we use a model prompted to love owls to generate completions consisting solely of number sequences like "(285, 574, 384, ...)". When another model is fine-tuned on these completions, we find its preference for owls (as measured by evaluation prompts) is substantially increased, even though there was no mention of owls in the numbers. This holds across multiple animals and trees we test. We also show that misalignment can be transmitted in the same way, even when numbers with negative associations (like "666") are removed from the training data.

Our experiment format is as follows. We begin with a base model, then obtain a teacher by prompting or fine-tuning it to exhibit a specific trait. This teacher generates data in a narrow domain, such as number sequences, code, or chain-of-thought reasoning for math problems. The data is filtered to remove any explicit references to the trait. Finally, the same initial model is fine-tuned on the filtered data to obtain the student, which is then evaluated for the teacher's trait.

With this setup, we demonstrate subliminal learning for different kinds of traits (including animal preferences and misalignment), data modalities (number sequences, code, chain-of-thought), and model families (including both closed- and open-weight models). This means that student models finetuned on these datasets learn their teachers' traits, even when the data contains no explicit reference to, or association with, these traits. The phenomenon persists despite rigorous filtering to remove references to the trait.