SoylentNews

Freeman writes:

https://arstechnica.com/ai/2025/06/why-one-man-is-archiving-human-made-content-from-before-the-ai-explosion/

Former Cloudflare executive John Graham-Cumming recently announced that he launched a website, lowbackgroundsteel.ai, that treats pre-AI, human-created content like a precious commodity—a time capsule of organic creative expression from a time before machines joined the conversation. "The idea is to point to sources of text, images and video that were created prior to the explosion of AI-generated content," Graham-Cumming wrote on his blog last week. The reason? To preserve what made non-AI media uniquely human.

[...] ChatGPT in particular triggered an avalanche of AI-generated text across the web, forcing at least one research project to shut down entirely.

That casualty was wordfreq, a Python library created by researcher Robyn Speer that tracked word frequency usage across more than 40 languages by analyzing millions of sources, including Wikipedia, movie subtitles, news articles, and social media. The tool was widely used by academics and developers to study how language evolves and to build natural language processing applications. The project announced in September 2024 that it will no longer be updated because "the Web at large is full of slop generated by large language models, written by no one to communicate nothing."

[...] The website points to several major archives of pre-AI content, including a Wikipedia dump from August 2022 (before ChatGPT's November 2022 release), Project Gutenberg's collection of public domain books, the Library of Congress photo archive, and GitHub's Arctic Code Vault—a snapshot of open source code buried in a former coal mine near the North Pole in February 2020. The wordfreq project appears on the list as well, flash-frozen from a time before AI contamination made its methodology untenable.

[...] As atmospheric nuclear testing ended and background radiation returned to natural levels, low-background steel eventually became unnecessary for most uses. Whether pre-AI content will follow a similar trajectory remains a question.

Still, it feels reasonable to protect sources of human creativity now, including archival ones, because these repositories may become useful in ways that few appreciate at the moment.

[...] For now, lowbackgroundsteel.ai stands as a modest catalog of human expression from what may someday be seen as the last pre-AI era.

Original Submission

_-iNT13- writes:

Over a week ago iStories published a story alleging Vladimir Vedeneev and Roman Venediktov, who own GlobalNet, allow the FSB access to Telegram messages. GlobalNet assigns IP addresses for Telegram and maintains it's networking equipment, and also has/had contracts with a FSB "research computing center" that helped plan the invasion of Ukraine and developed tools to de-anonymize internet users; and a flagship state-owned nuclear research laboratory.

Telegram responded to the allegations the following day - and in doing so made some glaring contradictions with what they told the court in Miami, USA.

More recently Meduza published a detailed story about both those iStories reports - the highlight being the revelation that "as part of a criminal case opened by Russia's Federal Security Service (FSB) in the spring of 2022, security agents have been intercepting Telegram messages."

Security researcher Michał "rysiek" Woźniak published a detailed report titled Telegram is indistinguishable from an FSB honeypot, which provides more information on how Telegram has been compromised.

Original Submission

Processed by jelizondo

Arthur T Knackerbracket has processed the following story:

Gaze into the temporal distance and you might spot the end of the age of silicon looming somewhere out there, as a research team at Penn State University claims to have built the first working CMOS computer entirely from two-dimensional materials.

The team, led by Pennsylvania State University engineering science professor Saptarshi Das, published a paper last week detailing the design and construction of their 2D one instruction set computer (OISC) based on the same complementary metal-oxide-semiconductor (CMOS) design that's a standard part of modern silicon-based computers. OISC is a minimalist abstract machine model that performs all operations using a single, universal instruction.

Does that mean we can expect to live through a post-silicon, 2D computing revolution? It won't be quite like that, Das told us. Rather, 2D CMOS computers will have specialized uses.

"They could become competitive in specialized domains such as edge AI, neuromorphic systems, or flexible electronics," Das told us.

The 2D machine they built is silicon-free, using molybdenum disulfide for n-type and tungsten diselenide for p-type transistors. The material pair "offer complementary electrical characteristics, relatively high mobility, and have demonstrated scalable growth via metal–organic chemical vapor deposition (MOCVD)," Das told The Register in an email. MOCVD was used to fabricate the team's 2D CMOS platform on sapphire wafers, with transistor channels just one atom thick. 

CMOS systems need both n- and p-type transistors (which move electrons along a circuit by having an excess and deficiency of electrons, respectively) to achieve the goal of CMOS computing - energy efficiency and reusability. That's why the team's 2D CMOS design is such a breakthrough, according to Das.

"We have demonstrated, for the first time, a CMOS computer built entirely from 2D materials," Das said in an announcement on the Penn State website.

[...] "[Scalability] is one of the most critical aspects of our work," Das told us. "While some steps (e.g., layer alignment and transfer) are still manual, most of the process is compatible with industry tools and can be automated."

Also see the Penn State press release

Journal Reference: Ghosh, S., Zheng, Y., Rafiq, M. et al. A complementary two-dimensional material-based one instruction set computer. Nature 642, 327–335 (2025). https://doi.org/10.1038/s41586-025-08963-7

Original Submission

looorg writes:

Your brain on ChatGPT: Accumulation of cognitive debt.

The more you rely on ChatGPT (and similar) things to do work for you the less work your brain does and eventually it becomes slow and lazy.

A new MIT study suggests that using AI writing assistants like ChatGPT can lead to what researchers call "cognitive debt" - a state where outsourcing mental effort weakens learning and critical thinking. The findings raise important questions about how large language models (LLMs) shape our brains and writing skills, especially in education.

The experiment divided 54 mostly college students from five Boston-area universities into three groups: one used OpenAI's GPT-4o (LLM group), another used traditional search engines (but no AI-generated answers), and a third wrote essays without any outside help (brain-only group)

It's a somewhat small study with only 54 participants. But if the trend is clear then the trend is clear and just finding more participants might not shift things. Also these are people attending universities or colleges so they should already know how to write papers.

The more you use tools to do brain work the weaker the brain got. After all the body is a lazy beast that likes to conserve energy.

Essentially, with AI, the brain worked less deeply and handed off more of the cognitive load to the tool.

One of the sharpest differences came in post-session interviews. After the first session, over 80% of LLM users struggled to accurately recall a quote from their just-written essay - none managed it perfectly.

That isn't hard to understand considering they didn't write it.

https://time.com/7295195/ai-chatgpt-google-learning-school/

https://the-decoder.com/mit-study-shows-cognitive-debt-through-chatgpt-heres-what-it-means-in-real-world-practice/

paper:
https://arxiv.org/pdf/2506.08872v1

Original Submission

Arthur T Knackerbracket has processed the following story:

The long-running battle of Germany's northernmost state, Schleswig-Holstein, to make a complete switch from Microsoft software to open-source alternatives looks close to an end. Many government operatives will permanently wave goodbye to the likes of Teams, Word, Excel, and Outlook in the next three months in a move to ensure independence, sustainability, and security.

Plans to go open-source were drawn up by Schleswig-Holstein as far back as 2017. In 2021, the state found another incentive to make the switch: Windows 11's hardware requirements. A move to LibreOffice and other open-source programs had a deadline of 2026 – there was no date set for ditching Windows at the time.

Last year brought news of a plan by the state to replace Windows with Linux and further expansion of LibreOffice, Open-Xchange, Nextcloud, and the Univention Active Directory (AD).

Now, Digitalisation Minister Dirk Schroedter has announced that "We're done with Teams!"

Question! Why should local governments use taxpayers' money to buy proprietary software from a single vendor? And what happens to citizens' data? A solution is to move to free software like Linux and LibreOffice – which is what Schleswig-Holstein is doing: https://t.co/P7cQJwEP7u pic.twitter.com/OuIHPlSteV

– LibreOffice (@LibreOffice)

Of the state's approximately 60,000 public servants, which includes civil servants, judges, and police officers, around half are transitioning away from Microsoft in this initial phase, with 30,000 more – mostly teachers – doing the same over the next few years.

Schroedter highlighted several reasons for the move. Money is obviously a major factor, with Microsoft's enterprise licensing fees reaching into the millions of euros.

[...] Other public bodies across the world are also moving away from Microsoft's products in favor of open-source or home-grown alternatives, from French police to India's defence ministry, writes France 24. Local governments in Denmark are also looking to ditch the firm.

Munich, the capital of the German state Bavaria, switched from Windows to Linux-based LiMux in 2004, though it switched back in 2017 as part of an IT overhaul. Wanting Microsoft to move its headquarters to Munich likely played a part in the return to Windows, too.

Original Submission

Arthur T Knackerbracket has processed the following story:

The percentage of women in the semiconductor industry is stubbornly low. According to a report released in April, 51 percent of companies report having less than 20 percent of their technical roles filled by women. At the same time, fewer of these companies were publicly committed to equal opportunity measures in 2024 than the year prior, the same report found.

This lack of support comes at the same time that major workforce shortages are expected, says Andrea Mohamed, COO and co-founder of QuantumBloom, which helps companies attract, retain, and advance early career women in STEM. The company focuses on the transition from higher education to the workforce, a critical point during which many women leave STEM.

[...] I understand the pressures that companies are facing around anything that's related to DEI. We need to change the conversation from DEI to talent management. This is retention and avoiding turnover costs. This is about needing every available brilliant mind in the United States that wants to be in semiconductors. We have offshored this industry for so long. Other countries have existing talent bases. We have to build it.

So there are semantics in all of this, but it's not just relabeling. This is about business. You are not going to be able to compete on a global stage in the United States if you are not finding ways to attract and retain new communities of workers, and women are one of those communities. That means understanding what women need from their employer, because if you do not provide it, they will go somewhere else that does. The concern by companies about, if they run a program like QuantumBloom, does that create a risk? It's the wrong question about risk. Your big risk is that your fab is empty, because you can't find workers and retain them.

In other industries, there are organizations that are very intentional about attracting and retaining their youngest talent. They are dedicating resources to investing in them, which is very rare - most organizations invest more the higher up you go. Really, we need to be thinking about flipping that script and investing more sooner.

upstart writes:

Starship 36 was preparing for 10th test flight from Texas when it underwent 'catastrophic failure' while on stand:

Starship 36 was preparing for 10th test flight from Texas when it underwent 'catastrophic failure' while on stand

[...] SpaceX said the rocket was preparing for the 10th flight test when it "experienced a major anomaly while on a test stand at Starbase", without elaborating on the nature of the complication.

[...] The Starship explosion occurred during a "routine static fire test", according to the Cameron County authorities.

During a static fire, part of the procedures preceding a launch, the Starship's Super Heavy booster would be anchored to the ground to prevent it from lifting off during the test firing.

[...] At 123 metres high (403ft), Starship is the world's largest and most powerful rocket and central to Musk's long-term vision of colonising Mars.

The Starship is billed as a fully reusable rocket with a payload capacity of up to 150 metric tonnes.

[...] The previous two outings also ended poorly, with the upper stage disintegrating over the Caribbean.

The failures will probably do little to dent Musk's space ambitions.

SpaceX has been betting that its "fail fast, learn fast" ethos, which has helped it dominate commercial spaceflight, will eventually pay off.

The company has caught the Super Heavy booster in the launch tower's giant robotic arms three times – a daring engineering feat it sees as being key to rapid reusability and slashing costs.

Nasa is increasingly relying on SpaceX, whose Dragon spacecraft is used to ferry astronauts to and from the International Space Station.

Original Submission

Arthur T Knackerbracket has processed the following story:

The LibreOffice project is preparing to cut some Windows support - and encourages users to switch to Linux.

The Document Foundation, the organization that backs and guides development of LibreOffice since Oracle dropped the ball, has a strong point of view about the future. Some of it is very visible, in a blog post about the looming end of Windows 10, but some is buried in the development notes about the work-in-progress next version, which will be LibreOffice 25.8.

The title of the blog post is self-explanatory. The end of Windows 10 is approaching, so it's time to consider Linux and LibreOffice. They're right. The post also links to the KDE-backed End of 10 campaign, which we covered a month ago. That site's list of places to go for help has been improved with a zoomable world map, but we feel it still badly needs some kind of hierarchical organization.

The blog also links to Distro Chooser, which is a noble idea, but with flawed execution. This site leads you through a set of questions and then recommends what Linux distributions might suit you. We found the presentation of the results overwhelming, though. It seems to be a list of all the candidates, color-coded according to how good a match they are. The Reg FOSS desk generally feels that "Less is more," and here, just distilling the results down to, say, a top three would far be more helpful.

We are sure that some people will dismiss any and all Linux distros as being inferior, just as they do of LibreOffice itself. That's not the point. The point is that it is a free alternative. You get the reward of breaking free of paid-for software you don't own.

Oracle washed its hands of OpenOffice and then dumped it on the Apache Foundation more than a decade ago.

OpenOffice does officially still exist but there hasn't been a new release in a couple of years, and we recommend avoiding it. LibreOffice is a direct in-place upgrade and will open the same files.

PiMuNu writes:

From:
https://www.bbc.co.uk/news/articles/cq8zxx9kk0ko

Two breakthrough Alzheimer's drugs have been deemed far too expensive, for too little benefit, to be offered on the NHS. The medicines are the first to slow the disease, which may give people extra time living independently.

The National Institute for Health and Care Excellence (NICE) concluded they were a poor use of taxpayers' money and said funding them could lead to other services being cut. Campaigners say it is a disappointment, but other dementia experts have also supported the decision.

The two drugs, donanemab and lecanemab, both help the body clear a gungy protein that builds up in the brains of people with Alzheimer's disease.

[...] The official price in the US is £20,000-£25,000 per patient per year. What the NHS would pay is confidential.

Around 70,000 people in England with mild dementia would have been eligible, potentially putting the bill in the region of £1.5bn a year for the drugs alone.

NHS resources, including infusing the drugs every two-to-four weeks and frequent brain scans to manage dangerous side effects, would also massively ramp up the cost.

Original Submission

fliptop writes:

From New York to California, state renewable electrical power dreams are collapsing. Power demands soar, while the federal government cuts funding and support for wind, solar, and grid batteries. Renewables cannot provide enough power to support the artificial intelligence revolution. The Net Zero electricity transition is failing in the United States:

For the last two decades, state governments have embraced policies aimed at replacing coal and natural gas power plants with renewable sources. Twenty-three states enacted laws or executive orders to move to 100% Net Zero electricity by 2050. Onshore and offshore wind, utility-scale and rooftop solar, and grid-scale batteries were heavily promoted by states and most federal administrations.

The New York State Climate Action Scoping Plan [PDF] of 2022 called for 70% renewable electricity by 2030 and 100% by 2040. But 49.7% of the state's electricity came from gas in 2024, up from 47.7% in 2023. A January executive order issued by President Trump halted federal leases for construction of offshore wind systems. New York, nine other east coast states, and California were counting on offshore wind in efforts to get to 100% renewable electricity, but new offshore wind projects are now halted.

Wind and solar have benefited from federal tax credits, loans, and outright grants since 1992. But the Trump administration is now working to slash federal government support for these technologies. The One Big Beautiful Bill Act (OBBB) passed the House of Representatives on May 22. The bill eliminates Production Tax Credits and Investment Tax Credits for renewable systems that begin construction later than 60 days after passage of the bill or for projects that do not complete construction by year end 2028. The bill also halts the sale of tax credits from renewable projects. If the Senate passes the bill, these measures will choke off green energy projects that have relied on federal funding for decades.

[...] Along with federal cutbacks, the artificial intelligence (AI) revolution now drives the nation's power system, interrupting the renewable electricity transition. Microsoft, Meta, Google, Amazon, and other giant firms are building new data centers and upgrading existing data centers to power AI. AI processors run 24-hours a day for months to enable computers to think like humans. When servers are upgraded to support AI, they consume 6 to 10 times more power than when used for cloud storage and the internet. Data centers consumed 4% of US electricity at the start of 2024 but are projected to consume 20% within the next decade.

[...] In December, the North American Electric Reliability Corporation concluded [PDF] that that over half of North America risks power shortfalls in the next decade from surging demand and coal and gas plant retirements. Grid operators are now stepping back from the transition to wind and solar. Coal-fired power plant closures have been postponed in Georgia, Indiana, Illinois, Tennessee, Utah, West Virginia, and other states. Nuclear plants are being restarted in Michigan and Pennsylvania. But the big winner will be natural gas.

More than 200 gas plants are planned or under construction. Gas facilities can be brought online in about three years, compared to ten years for nuclear plants. Gas plants can be built near cities, often on former power plant sites, and require fewer new transmission lines than needed by wind and solar systems.

TFA talks about BYOP (bring your own power) trends for AI and the perils of using lithium batteries for storing wind and solar generated energy.

Related:

• We did the Math on AI's Energy Footprint. Here's the Story you Haven't Heard.
• AWS Claims That Britain Needs More Nuclear Power for AI Datacenters
• PA's Largest Coal Plant to Become 4.5GW Gas-Fired AI Hub

Original Submission

PiMuNu writes:

Physics reports that Fermilab's g-2 collaboration has now published its final result.

https://physics.aps.org/articles/v18/116

After measuring the wobbles of 300 billion muons, the Muon g − 2 Collaboration has pinpointed with exquisite precision the internal magnetism of these subatomic particles. The muon's magnetic strength, or moment, has animated particle physics research over the past two decades, as experiment and theory appeared to disagree over its value—raising a flag for possible new physics. In a somewhat surprising turn of events, the final results from the Muon g − 2 experiment line up with the most recent predictions, further validating the standard model of particle physics.

The muon—the heavy cousin to the electron—started to grab the particle-physics spotlight in the 1990s when an experiment at Brookhaven National Laboratory in New York reported the first hints that the muon's magnetic behavior might not match predictions based on the standard model, which has otherwise been widely successful in explaining the subatomic world. The Brookhaven measurements involved magnetically trapping muons in a circular ring and observing how much their internal magnet, or "spin," wobbled around an applied magnetic field. To further investigate this discrepancy, the experiment's big magnet was moved cross-country in 2013 to Fermi National Laboratory (Fermilab) in Illinois. The first results from the transplanted Muon g − 2 experiment came out in 2021, showing good agreement with the Brookhaven findings and raising the significance of the discrepancy (see Viewpoint: Muon's Escalating Challenge to the Standard Model).

The discrepancy has apparently been resolved following improvements to the theoretical estimate for g-2. The Standard Model of Particle Physics - disappointingly for particle physicists - still stands as an entirely accurate prediction of almost all of subatomic physics (the exception being neutrino oscillations https://en.wikipedia.org/wiki/Neutrino_oscillation ).

Preprint paper:
http://arxiv.org/abs/2506.03069

Previously:
    • g-2 Releases New Data
    • Making Sense of the Muon's Misdemeanours
    • Muon G-2 Experiment Hints at Mysterious New Physics
    • Latest Muon Measurements Hint at Cracks in the Standard Model
    • The Cloak-and-Dagger Tale Behind This Year's Most Anticipated Result in Particle Physics
    • Precision Measurements of the Standard Model

Original Submission

cereal_burpist writes:

Welcome to the 'infinite workday' of 8 p.m. meetings and constant messages

While probably not news to many Soylentils, it's interesting to me that this comes from (a Big-Bad-Corp like) Microsoft. Will they, or any companies, implement changes based on these findings? [endless laughter]

Workers are struggling to cope with a "seemingly infinite workday," involving an increasing load of meetings scheduled at 8 p.m. or later and a near-constant stream of interruptions, according to new research by Microsoft. The company analyzed data from users of Microsoft 365 services... globally between mid-January and mid-February. It found that the number of meetings booked between 8 p.m. and just before midnight had risen 16% compared with last year. Geographically dispersed teams, as well as those with flexible working arrangements, were responsible for much of that increase.

"The infinite workday... starts early, mostly in email, and quickly swells to a focus-sapping flood of messages, meetings, and interruptions," Microsoft said in a report Tuesday.

The company found that the average worker is interrupted every two minutes by a meeting, an email or a chat notification during a standard eight-hour shift — adding up to 275 times a day.

And those messages don't stop after they've clocked off. During the study period, the average employee sent or received 58 instant messages outside of their core working hours — a jump of 15% from last year.

The typical worker also receives 117 emails per day and, by 10 p.m., almost one-third of employees are back in their inboxes, "pointing to a steady rise in after-hours activity," Microsoft noted.

[...] "It's the professional equivalent of needing to assemble a bike before every ride. Too much energy is spent organizing chaos before meaningful work can begin," it added.

[...] One outcome is that one-third of workers feel it has been "impossible to keep up" with the pace of work over the past five years, according to a Microsoft-commissioned survey of 31,000 employees around the world, cited in the Tuesday report.

Original Submission

An Anonymous Coward writes:

https://www.wsj.com/health/wellness/marijuana-to-treat-autism-some-parents-say-yes-1d9a3ad1
https://archive.ph/ZK2KH

Marijuana is becoming easier to get in many states, and one group showing interest might surprise you: parents of children with autism.
In online and support groups, families swap tips and share experiences—even though the science is still inconclusive. Most doctors don't prescribe cannabis and usually advise against it.

But the few who do say demand is rising. Dr. Mohsin Maqbool, a pediatric neurologist in Plano, Texas, says about a third of his patients are children with autism and about 40% of them treated with cannabis.

He prescribes a combination of THC, the main psychoactive component of marijuana, and CBD, another part of the cannabis plant that doesn't produce a high and can counteract some of the impact of THC.

While medical marijuana is prescribed for everything from cancer symptoms to epilepsy to chronic pain and even dementia, its use in children with autism is more controversial.

After all, we know regular use of THC in adolescents can negatively impact their developing brains and lead to a higher risk of mental-health problems.

But recreational and medical use of marijuana are quite different. These patients aren't getting high—their families and doctors are intentionally avoiding that by strictly regulating and individualizing the dose.
And the parents of children with autism are often desperate to try something to help treat their children for a neurodevelopmental condition that has no definitive cause. They say cannabis is no more risky than the other medications they've tried unsuccessfully, often with many side effects.

Take Marlo Jeffrey, a nurse practitioner who lives outside Dallas. She says she has tried countless medications for her son, Jaiden Gaut. He is an 11-year-old who was diagnosed with ADHD and high-functioning autism when he was 5. His symptoms include anger and aggression, sometimes escalating to self-harm and violence.

Jeffrey took her son to see Dr. Maqbool last fall. Jaiden started THC/CBD gummies, escalated his dose over a month, until he got up to 7.5 milligrams of THC and CBD each a day.

"His behavior improved a lot during that time," says Jeffrey. Jaiden was able to stop taking another medication that had caused him to gain a lot of weight. His physical aggression and cursing improved. He stopped using repetitive words as much and his thinking became more flexible. Since then, he has transitioned to a mainstream school and his grades have gone up.

But anecdotal success doesn't equate to scientific validation. Many doctors aren't yet convinced cannabis is a safe and effective treatment. Groups like the American Academy of Child & Adolescent Psychiatry warn against treating autistic children with cannabis. The Autism Science Foundation says parents should be "very cautious about giving THC to their children" and more research needs to be done on CBD. Some researchers and doctors say CBD shows promising signs as a treatment for autism but are wary of THC.

Most CBD consists of trace amounts of THC. Dr. Doris Trauner, neuroscience professor at the UC San Diego School of Medicine, says she has had patients whose CBD from dispensaries had higher-than-expected concentrations of THC, and they ended up experiencing worse behavior. "Children became aggressive, some almost psychotic," she says. "That's not across the board, of course. But I do think that a lot of physicians would be concerned about using high amounts of THC."

owl writes:

https://arstechnica.com/gadgets/2025/06/reddit-user-surprised-when-1960s-computer-panel-emerged-from-collapsed-family-garage/

Recently, a Reddit user discovered a rare RCA Spectra 70/35 computer control panel from 1966 in their family's old collapsed garage, posting photos of the pre-moon landing mainframe component to the "retrobattlestations" subreddit that celebrates vintage computers. After cleaning the panel and fixing most keyswitches, the original poster noted that actually running it would require "1,500lbs of mainframe"—the rest of the computer system that's missing.

As it turns out, the panel had been sitting in the garage for decades without the poster's knowledge. "In short my house is a two-family, my dad used to rent out the other half before I was born," explained SonOfADeadMeme in the thread on Friday. "One of the people who rented out the apartment worked at IBM (apparently the RCA Spectra 70's were compatible with IBM sets from the time) and shortly before he left he shown [sic] up with a forklift and left something in the garage."

Original Submission

upstart writes:

Zero-Click AI Vulnerability Exposes Microsoft 365 Copilot Data Without User Interaction:

A novel attack technique named EchoLeak has been characterized as a "zero-click" artificial intelligence (AI) vulnerability that allows bad actors to exfiltrate sensitive data from Microsoft 365 Copilot's context sans any user interaction.

The critical-rated vulnerability has been assigned the CVE identifier CVE-2025-32711 (CVSS score: 9.3). It requires no customer action and has been already addressed by Microsoft. There is no evidence that the shortcoming was exploited maliciously in the wild.

"AI command injection in M365 Copilot allows an unauthorized attacker to disclose information over a network," the company said in an advisory released Wednesday. It has since been added to Microsoft's Patch Tuesday list for June 2025, taking the total number of fixed flaws to 68.

Aim Security, which discovered and reported the issue, said it's instance of a large language model (LLM) Scope Violation that paves the way for indirect prompt injection, leading to unintended behavior.

LLM Scope Violation occurs when an attacker's instructions embedded in untrusted content, e.g., an email sent from outside an organization, successfully tricks the AI system into accessing and processing privileged internal data without explicit user intent or interaction.

"The chains allow attackers to automatically exfiltrate sensitive and proprietary information from M365 Copilot context, without the user's awareness, or relying on any specific victim behavior," the Israeli cybersecurity company said. "The result is achieved despite M365 Copilot's interface being open only to organization employees."

In EchoLeak's case, the attacker embeds a malicious prompt payload inside markdown-formatted content, like an email, which is then parsed by the AI system's retrieval-augmented generation (RAG) engine. The payload silently triggers the LLM to extract and return private information from the user's current context.

[...] "The attack results in allowing the attacker to exfiltrate the most sensitive data from the current LLM context - and the LLM is being used against itself in making sure that the MOST sensitive data from the LLM context is being leaked, does not rely on specific user behavior, and can be executed both in single-turn conversations and multi-turn conversations."

EchoLeak is especially dangerous because it exploits how Copilot retrieves and ranks data – using internal document access privileges – which attackers can influence indirectly via payload prompts embedded in seemingly benign sources like meeting notes or email chains.

Original Submission