Deezer says AI-generated songs now make up 44% of all new uploads to its platform, with nearly 75,000 arriving each day and more than two million per month. The company notes that consumption of these tracks is still very low, "between 1-3% of the total streams," and 85% are flagged as fraudulent. TechCrunch reports: The latest figure from Deezer highlights a continuous surge in AI-generated music uploads to the platform. Deezer reported receiving around 60,000 AI tracks per day in January, up from 50,000 in November, 30,000 in September, and just 10,000 in January 2025, when it first launched its AI-music detection tool.

Songs tagged as AI-generated on Deezer are automatically removed from algorithmic recommendations and not included in editorial playlists. The company announced today that it will no longer store hi-res versions of AI tracks. "AI-generated music is now far from a marginal phenomenon and as daily deliveries keep increasing, we hope the whole music ecosystem will join us in taking action to help safeguard artists' rights and promote transparency for fans," said Deezer CEO Alexis Lanternier in a press release. "Thanks to our technology and the proactive measures we put in place more than a year ago, we have shown that it's possible to reduce AI-related fraud and payment dilution in streaming to a minimum."

DeanonymizedCoward writes: Engadget reports that Palantir has posted to X a summary of CEO Alex Karp and Nicholas W. Zamiska's 2025 book, The Technological Republic, which reads like a utopian idealist doodled on a Bond villain's whiteboard. While the post makes some decent points, it also highlights the Big-AI attitude that the AI surveillance state is in fact a good thing, and strongly implies that the Good Guys need to do war crimes before the Bad Guys get around to it. "The ability of free and democratic societies to prevail requires something more than moral appeal," one of the 22 points states. "It requires hard power, and hard power in this century will be built on software."

The book is billed as "a passionate call for the West to wake up to our new reality," and other excerpts in the social media post include assertions such as: "Free email is not enough. The decadence of a culture or civilization, and indeed its ruling class, will be forgiven only if that culture is capable of delivering economic growth and security for the public"; "National service should be a universal duty"; "The postwar neutering of Germany and Japan must be undone"; and "Some cultures have produced vital advances; others remain dysfunctional and regressive."

The statement criticizes the West's resistance to "defining national cultures in the name of inclusivity," as well as the treatment of billionaires and the "ruthless exposure of the private lives of public figures."

An anonymous reader quotes a report from Bloomberg, written by writer Austin Carr: Allbirds is pivoting to artificial intelligence. The San Francisco brand, whose wool running shoes were once the sneaker du jour among the tech crowd, announced last week that it was expanding into AI computing infrastructure. The bizarre strategic shift was immediately greeted with a surprising frenzy on Wall Street, where shares of Allbirds soared 582% last Wednesday before dropping the next day. [...] Of course, the absurdity of Allbirds' situation echoed familiar Silicon Valley tropes -- from the endless startup pivots of the 2010s to the more recent boom-and-bust cycles of arbitrarily valued crypto coins. But it immediately reminded me of the marketing ploys of the dot-com crash. After all, some of the more iconic fails ended up being retailers such as Pets.com, Webvan, etc., riding the web wave with little to show for it beyond terrible margins.

One particular comparison from that period stands out as relevant to Allbirds: Zap.com. The holding company behind it, Zapata Corp., had a long and convoluted history, but was essentially selling fish-oil products by the time it decided to reinvent itself as an internet portal. It amassed a variety of web properties -- in media, e-commerce, gaming and so on -- and even once tried to acquire the search engine Excite. Spoiler alert: Zap flopped. Jen Heck, then a young employee at one of Zap's up-and-coming portfolio entities, remembers how quickly the hype of that web 1.0 turned to hell. As absurd as Zapata's pivot sounds today, it seemed feasible during the excitement of the internet revolution. "We went from like, 'Wow, this life thing is just so easy,' to it all ending so suddenly," Heck recalls. The ones who survived that tech bubble, she says, actually had differentiated products and the right creative thinkers building them -- and weren't just cynically jumping on the latest hot trend. "'Internet' was the magic word then, and 'AI' is the magic word now," Heck says.

Axios reports that the NSA is using Anthropic's restricted Mythos Preview model despite the Pentagon insisting the company poses a "supply chain risk." Axios reports: The government's cybersecurity needs appear to be outweighing the Pentagon's feud with Anthropic. The department moved in February to cut off Anthropic and force its vendors to follow suit. That case is ongoing. The military is now broadening its use of Anthropic's tools while simultaneously arguing in court that using those tools threatens U.S. national security.

Two sources said the NSA was using Mythos, while one said the model was also being used more widely within the department. It's unclear how the NSA is currently using Mythos, but other organizations with access to the model are using it predominantly to scan their own environments for exploitable security vulnerabilities.

Anthropic restricted access to Mythos to around 40 organizations, contending that its offensive cyber capabilities were too dangerous to allow for a wider release. Anthropic only announced 12 of those organizations. One source said the NSA was among the unnamed agencies with access. The NSA's counterparts in the U.K. have said they have access to the model through the country's AI Security Institute. Anthropic's CEO met with top U.S. officials on Friday to discuss "opportunities for collaboration," according to a White House spokesperson, "as well as shared approaches and protocols to address the challenges associated with scaling this technology."

There's been a few complaints about Amazon's drone delivery service. "The automated mailmen are dropping off packages from 10 feet in the air," reports the New York Post, "rendering the contents of each box susceptible to crashing and smashing."

One example? Tamara Hancock filmed a drone delivering a bottle of Torani flavoring syrup to her home in Arizona (as a test of how Amazon handled fragile items). It was delivered it in a plastic bottle — not glass — but the massive drone drops the drone from so high that the impact cracked the bottle's cap. (In the video Hancock opens her delivery to find leaked flavoring syrup "everywhere.")

The delivery was hard to film, Hancock says, because "If the drone sees me in the back yard, it will not drop, because it is worried about hurting humans or animals." The Post notes Amazon's "AI-charged fleet" of drones are "Outfitted with industry-leading 'sense and avoid' technology, the aerodynamic machines are equipped to drop off eligible items, weighing a maximum of five pounds, at designated areas in 60 minutes or less." The high-tech, however, apparently does not ensure gentle landings. Collisions, including a recent crash-and-burn into a Texas building, as well as several mid-flight malfunctions in rainy weather, have abounded since the drones' inaugural launch....

Tasha, a separate Amazon user, spotted the drone plunging a package near the paved driveway of a neighbor's yard. Unfortunately, its propellers caused other, previously delivered parcels to blow away, sending one into the street... In a statement to The Post, Amazon said it apologized for one of the "rare instances when products don't arrive as expected."
Amazon's drone fleet has been running since late 2024, the Post adds, and are now offering "ultra-fast" shipping in U.S. states including Arizona, Florida, Michigan, Kansas and Texas.

The machines do seem massive. I'm surprised neighbors aren't complaining about the noise...

Zoom "has partnered with World, Sam Altman's iris-scanning identity company (previously known as Worldcoin), " reports Digital Trends, "to add real-time human verification inside meetings." Zoom is now inviting organizations to join the beta version of the rollout, which Digital Trends says "lets hosts confirm that every face on the call belongs to a real person, not an AI-generated imposter. " For those wondering how World's Deep Face technology works, it includes a three-step process. It cross-references a signed image from a user's original Orb registration, a live face scan from the device, and the frame of the video that's visible to the other participants in the meeting. Only when the three samples match does a "Verified Human" badge appear next to the user's name...

Hosts can also make Deep Face verification mandatory for joining meetings, preventing unverified participants from joining entirely. Mid-call, on-the-spot checks are also possible...

Live Science spoke with physicist David Gross, who today received the $3 million "Special Breakthrough Prize in Fundamental Physics". He was part of a trio that won the 2004 physics Nobel prize for research that helped complete the Standard Model of particle physics. But when asked if physics will reach a unified theory of the fundamental forces of nature within 50 years, Gross has a surprising answer. "Currently, I spend part of my time trying to tell people... that the chances of you living 50 [more] years are very small."

Cold War estimates for a 1% chance of nuclear war each year seem low, Gross says. "The chances are more likely 2%. So that's a 1-in-50 chance every year." David Gross: The expected lifetime, in the case of 2% [per year], is about 35 years. [The expected lifetime is the average time it would take to have had a nuclear war by then. It is calculated using similar equations as those used to determine the "half-life" of a radioactive material.]

Live Science: So what do you suggest as remedies to lower that risk?

Gross: We had something called the Nobel Laureate Assembly for reducing the risk of nuclear war in Chicago last year. There are steps, which are easy to take — for nations, I mean. For example, talk to each other. In the last 10 years, there are no treaties anymore. We're entering an incredible arms race.

We have three super nuclear powers. People are talking about using nuclear weapons; there's a major war going on in the middle of Europe; we're bombing Iran; India and Pakistan almost went to war. OK, so that's increased the chance [of nuclear war]. I would really like to have a solid estimate — it might be more, and I think I'm being conservative — but a 2% estimate [of nuclear war] in today's crazy world.

Live Science: Do you think we'll ever get to a place where we get rid of nuclear weapons?

Gross: We're not recommending that. That's idealistic, but yes, I hope so. Because if you don't, there's always some risk an AI 100 years from now [could launch nuclear weapons], but chances of [humanity] living, with this estimate, 100 years, is very small, and living 200 years is infinitesimal. So [the answer to] Fermi's question of "Where are the civilizations, all the intelligent organisms around the galaxy, and why don't they talk to us?" is that they've killed themselves...

There are now nine nuclear powers. Even three is infinitely more complicated than two. The agreements, the norms between countries, are all falling apart. Weapons are getting crazier. Automation, and perhaps even AI, will be in control of those instruments pretty soon... It's going to be very hard to resist making AI make decisions because it acts so fast.
He points out that with the threat of climate change, "people have done something," even though "It's a much harder argument to make than about nuclear weapons.

"We made them; we can stop them."

Thanks to hwstar (Slashdot reader #35,834) for sharing the article.

Last May Duolingo's stock peaked at $529.05. But while the learning app passed $1 billion in revenue in 2025 and 50 million daily active users, today its stock price has dropped more than 81%, to $100.51.

And there's been other changes, reports Entrepreneur: In April 2025, Duolingo CEO Luis von Ahn made headlines after writing a memo calling the company "AI-first." In the memo, von Ahn announced that the language-learning platform would track employees' AI use in performance reviews. Now, a year later, von Ahn is backtracking and rethinking how he measures employee performance. He told the Silicon Valley Girl podcast earlier this month that Duolingo no longer considers AI use in performance reviews.

The change arose after employees started to ask, "Do you just want us to use AI for AI's sake?" von Ahn explained. "We said no, look — the most important thing in your performance is that you are doing whatever your job is as well as possible. A lot of times, AI can help you with that, but if it can't, I'm not going to force you to do that," von Ahn said on the podcast. He felt as though the company was "trying to push something that in some cases did not fit" instead of "being held accountable for the actual outcome." The CEO is, however, still sticking to other "constructive constraints" he introduced in the April 2025 memo, including stopping contractor hiring in cases where AI can assume their workload...

Von Ahn also mentioned that a few months ago, Duolingo had a day dedicated to vibe coding, or prompting AI to create an app without manually writing a single line of code. Every single person at the company, from engineers to human resources professionals, had to vibe code an app. Vibe coding has made an impact at the company. One of Duolingo's latest offerings, a course teaching users how to play chess, arose when two people vibe-coded the first prototype of it, the CEO said. Neither of them knew how to play chess or program, but they managed to use AI to create the whole chess curriculum and a prototype of the app in about six months last year. Now chess is Duolingo's fastest-growing course, according to von Ahn. "At this point, we have seven million daily active users that are learning chess," the CEO said on the podcast.

"A trailer has been released for the first film to star an authorised generative AI version of a major Hollywood actor," writes The Guardian: Val Kilmer was cast in western As Deep As the Grave before his death in April 2025. Production delays meant he never shot any scenes, but the creative team worked with UK-based company Sonantic to create an AI speaking voice based on his old recordings. His estate and daughter Mercedes collaborated with the film-makers on the visual deepfake of the actor. Kilmer, who was diagnosed with throat cancer, was also assisted by technology for his cameo in 2022's Top Gun: Maverick...

Writer-director Coerte Voorhees confirmed that Kilmer is seen for around an hour of the film's running time... Voorhees has said that the production followed Sag-Aftra [union] guidelines, and that Kilmer's estate — which provided archival material for them to use — was compensated financially.
"Kilmer's likeness can be seen portraying Father Fintan, a Catholic priest and Native American spiritualist," adds The Hollywood Reporter. But the AV Club calls it "ghoulish puppet show time."

"Having your AI Val Kilmer puppet whisper 'Don't fear the dead, and don't fear me' in a movie trailer is a bold choice..." He is accompanied (per Variety) by a whole host of disclaimers, caveats, and explanations offered by writer-director Coerte Voorhees and his associates: Kilmer deeply wanted to be in the movie, but was too sick to do so. His family endorses and supports his inclusion. He was a big fan of technology, including, presumably, its use in turning his own image into a digital avatar to then shove into movies...

The fact is, of course, that nobody would be paying a fraction of this attention to As Deep As The Grave — about early female archeologist Ann Axtell Morris — if it weren't now being used as the stage on which Voorhees was very publicly accepting the dare to go full-on ghoulish with AI tech.
"The filmmakers said they hoped they were showing Hollywood how to use the technology in a positive way..." notes Australia's ABC News. But their articles add that "Some have called the trailer 'terrifying' and 'disgusting' on social media."

Mashable writes: "Very fitting that this trailer includes a scene where a corpse is unceremoniously yanked out of the ground," read one of the top comments on As Deep as the Grave's trailer at time of writing... [O]nline commenters have labelled it disgusting and disrespectful, not only for digitally reanimating Kilmer but also for the damaging precedent As Deep as the Grave's use of AI could set for the film industry as a whole.

"From 2008 to 2024, the number of four-year computer science degrees granted rose about fivefold..." reports the Washington Post. Then in 2025 CS suddenly dropped from the fourth-largest undergraduate major to sixth, they report (citing data from the nonprofit National Student Clearinghouse, which compiles numbers from 97% of U.S. universities.

The 54,000-student drop was "the biggest one-year drop of any major discipline going back to at least 2020." But what major are they choosing instead? Sarah Karamarkovich, a research associate with the National Student Clearinghouse, pointed to an explanation from the data that we had overlooked. Enrollments in two interdisciplinary majors, data analytics and data science, topped a combined 35,000 in the fall of 2025. That was up from a few hundred when those disciplines were broken out into their own majors in 2020. Those relatively new categories reflect colleges' zeal to create specialized majors, including in AI, data science, robotics and cybersecurity. Some of those disciplines may be counted in the national enrollment data as computer science. Others are not.

The numbers suggest that some of the disappearing computer science majors didn't flee so much as they splintered into related disciplines.... The 8 percent decline in computer science majors last fall was nearly mirrored by a 7.3 percent increase in engineering majors, according to the National Student Clearinghouse data. Within engineering, mechanical and electrical engineering major enrollments increased by the largest absolute amounts — a jump of 11 percent and 14 percent, respectively.

Friday Anthropic's CEO met with top U.S. officials and "discussed opportunities for collaboration," according to a White House spokesperson itedd by Politico, "as well as shared approaches and protocols to address the challenges associated with scaling this technology."

CNN notes the meeting happens at the same time Anthropic "battles the Trump administration in court for blacklisting its Claude AI model..." The meeting took place as the US government is trying to balance its hardline approach to Anthropic with the national security implications of turning its back on the company's breakthrough technology — including its Mythos tool that can identify cybersecurity threats but also present a roadmap for hackers to attack companies or the government... The Office of Management and Budget has already told agencies it is preparing to give them access to Mythos to prepare, Bloomberg reported. Axios reported the White House is also in discussion to gain access to Mythos.
The Trump administration "recognizes the power" of Mythos, reports Axios, "and its highly sophisticated — and potentially dangerous — ability to breach cybersecurity defenses." "It would be grossly irresponsible for the U.S. government to deprive itself of the technological leaps that the new model presents," a source close to negotiations told us. "It would be a gift to China"... Some parts of the U.S. intelligence community, plus the Cybersecurity and Infrastructure Security Agency (CISA, part of Homeland Security), are testing Mythos. Treasury and others want it.
The White House added they plan to invite other AI companies for similar discussions, Politico reports. But Mythos "is also alarming regulators in Europe, who have told POLITICO they have not been able to gain access..." U.S. government agency tech leaders sought access to the model after Anthropic earlier this year began testing the model and granted limited access to a select group of companies, including JPMorgan, Amazon and Apple... after finding it had hacking capabilities far outstripping those of previous AI models. This includes the ability to autonomously identify and exploit complex software vulnerabilities, such as so-called zero-day flaws, which even some of the sharpest human minds are unable to patch. The AI startup also wrote that the model could carry out end-to-end cyberattacks autonomously, including by navigating enterprise IT systems and chaining together exploits. It could also act as a force-multiplier for research needed to build chemical and biological weapons, and in certain instances, made efforts to cover its tracks when attacking systems, according to Anthropic's report on the model's capabilities and its safety assessments.

Those findings and others have inspired fears that the model could be co-opted to launch powerful cyberattacks with relative ease if it fell into the wrong hands. Logan Graham, a senior security researcher at Anthropic, previously told POLITICO that researchers and tech firms had been given early access to Mythos so they could find flaws in their critical code before state-backed hackers or cybercriminals could exploit them. "Within six, 12 or 24 months, these kinds of capabilities could be just broadly available to everybody in the world," Graham said.

Some failed startups are reportedly selling old Slack messages, emails, and other internal records to AI companies as training data, creating a new way to cash out after shutting down. Fast Company reports: Shanna Johnson, the CEO of now-defunct software company Cielo24, told the publication that she was able to sell every Slack message, internal email, and Jira ticket as training data for "hundreds of thousands of dollars."

This isn't a one-off scenario. SimpleClosure, a startup that helps companies like Cielo24 shut down, told Forbes that there's been major interest from AI companies trying to get their hands on workplace data. Because of this, SimpleClosure launched a new tool that allows companies to sell their wealth of internal communications -- from Slack archives to email chains -- to AI labs. The company said it's processed 100 such deals in the past year. Payouts ranged from $10,000 to $100,000. "I think the privacy issues here are quite substantial," Marc Rotenberg, founder of the Center for AI and Digital Policy, told Forbes. "Employee privacy remains a key concern, particularly because people have become so dependent on these new internal messaging tools like Slack. ... It's not generic data. It's identifiable people."

BrianFagioli writes: Mozilla's email subsidiary MZLA Technologies just introduced Thunderbolt, an open-source AI client aimed at organizations that want to run AI on their own infrastructure instead of relying entirely on cloud services. The idea is to give companies full control over their data, models, and workflows while still offering things like chat, research tools, automation, and integration with enterprise systems through the Haystack AI framework. Native apps are planned for Windows, macOS, Linux, iOS, and Android. Thunderbolt allows organizations to do the following:
- Run AI with their choice of models, from leading commercial providers to open-source and local models
- Connect to systems and data: Integrate with pipelines and open protocols, including: deepset's Haystack platform, Model Context Protocol (MCP) servers, and agents with the Agent Client Protocol (ACP)
- Automate workflows and recurring tasks: Generate daily briefings, monitor topics, compile reports, or trigger actions based on events and schedules
- Work seamlessly across devices with native applications for Windows, macOS, Linux, iOS, and Android
- Maintain security with self-hosted deployment, optional end-to-end encryption, and device-level access controls

An anonymous reader quotes a report from the Wall Street Journal: An agreement with the Philippines to establish a high-tech industrial hub is the Trump administration's latest effort to lessen China's dominance over global supply chains. The deal to build up American manufacturing across a stretch of the island of Luzon, signed Thursday, will offer U.S. companies access to essential inputs such as critical minerals that bypass Beijing's control. The artificial-intelligence-powered manufacturing hub is planned for a 4,000-acre site given to the U.S. by Manila, said undersecretary of State for Economic Affairs Jacob Helberg. The U.S. will occupy the site rent-free and administer it as a special economic zone.

The hub will have diplomatic immunity, such as the protections afforded to an American embassy, and operate under U.S. common law -- the first arrangement of its kind anywhere in the world. The two-year lease is renewable for 99 years. [...] "You can't build anything in Ohio if the minerals and the process materials are controlled by an adversary who can cut you off tomorrow," Helberg said in an interview. [...] The planned manufacturing hub is largely conceptual at this stage, and details, including which American companies will participate and just what they will build in the Philippines, are yet to be determined.

[...] The administration will ask companies to put forward proposals to compete for a spot in building out the hub, giving priority to bids that will help move critical minerals processing and manufacturing off Chinese suppliers. Investment will have to come from private-sector companies -- not the U.S. government. Factories approved for operation in the hub will be highly automated, Helberg said, using autonomous systems to operate around the clock. The Philippines has a history of robust manufacturing, particularly in semiconductors, but that has stagnated in recent decades because of high energy and logistics costs. Companies will have to address in their proposals how they will contend with energy costs and workforce needs; they can send American workers overseas or hire locally, Helberg said.

Intel has launched a new budget-focused Core Series 3 processor line for lower-cost laptops -- "Intel's response to budget CPUs that are appearing in laptops like the Apple MacBook Neo," writes PCWorld's Mark Hachman. From the report: Intel unexpectedly launched the Core Series 3, based on its excellent "Panther Lake" (Core Ultra Series 3) architecture and 18A manufacturing, for devices for home consumers and small business on Thursday. Intel announced that a number of partners will launch laptops based upon the chip, including Acer, Asus, HP, Lenovo, and others. Although those laptops will be available beginning today, a number of them will begin shipping later this year, the partners said.

All of it -- from the specifications down to the messaging -- feels extremely aimed at trimming the fat and delivering to users just what they'll want. Intel's new Core Series 3 family just includes two "Cougar Cove" performance cores and four low-power efficiency "Darkmont" cores, with two Xe graphics cores on top of it. Intel isn't really worrying about AI, with an NPU capable of just 17 TOPS, though the company claims the CPU, NPU, and GPU combined reach 40 TOPS of performance. Yes, laptops will use pricey DDR5 memory, but at the lower end: just DDR5-6400 speeds. Support for three external displays will be included, though, maximizing multiple screens for maximum productivity. Intel used the term "all day battery life" without elaboration.

[...] Intel Core Series 3 delivers up to 47 percent better single-thread performance, up to 41 percent better multi thread performance, and up to 2.8x better GPU AI performance, Intel said. Compared against Intel's older Core 7 150U, Intel is saying that the new chip will outperform it by 2.1 times in content-creation and 2.7 times the AI performance. [...] We still don't know what Intel will charge for the chip, nor do we know what you'll be able to buy a Core Series 3 laptop for.

An anonymous reader quotes a report from Ars Technica: Two years ago, Microsoft launched its first wave of "Copilot+" Windows PCs with a handful of exclusive features that could take advantage of the neural processing unit (NPU) hardware being built into newer laptop processors. These NPUs could enable AI and machine learning features that could run locally rather than in someone's cloud, theoretically enhancing security and privacy. One of the first Copilot+ features was Recall, a feature that promised to track all your PC usage via screenshot to help you remember your past activity. But as originally implemented, Recall was neither private nor secure; the feature stored its screenshots plus a giant database of all user activity in totally unencrypted files on the user's disk, making it trivial for anyone with remote or local access to grab days, weeks, or even months of sensitive data, depending on the age of the user's Recall database.

After journalists and security researchers discovered and detailed these flaws, Microsoft delayed the Recall rollout by almost a year and substantially overhauled its security. All locally stored data would now be encrypted and viewable only with Windows Hello authentication; the feature now did a better job detecting and excluding sensitive information, including financial information, from its database; and Recall would be turned off by default, rather than enabled on every PC that supported it. The reconstituted Recall was a big improvement, but having a feature that records the vast majority of your PC usage is still a security and privacy risk. Security researcher Alexander Hagenah was the author of the original "TotalRecall" tool that made it trivially simple to grab the Recall information on any Windows PC, and an updated "TotalRecall Reloaded" version exposes what Hagenah believes are additional vulnerabilities.

The problem, as detailed by Hagenah on the TotalRecall GitHub page, isn't with the security around the Recall database, which he calls "rock solid." The problem is that, once the user has authenticated, the system passes Recall data to another system process called AIXHost.exe, and that process doesn't benefit from the same security protections as the rest of Recall. "The vault is solid," Hagenah writes. "The delivery truck is not." The TotalRecall Reloaded tool uses an executable file to inject a DLL file into AIXHost.exe, something that can be done without administrator privileges. It then waits in the background for the user to open Recall and authenticate using Windows Hello. Once this is done, the tool can intercept screenshots, OCR'd text, and other metadata that Recall sends to the AIXHost.exe process, which can continue even after the user closes their Recall session.

"The VBS enclave won't decrypt anything without Windows Hello," Hagenah writes. "The tool doesn't bypass that. It makes the user do it, silently rides along when the user does it, or waits for the user to do it." A handful of tasks, including grabbing the most recent Recall screenshot, capturing select metadata about the Recall database, and deleting the user's entire Recall database, can be done with no Windows Hello authentication. Once authenticated, Hagenah says the TotalRecall Reloaded tool can access both new information recorded to the Recall database as well as data Recall has previously recorded. "We appreciate Alexander Hagenah for identifying and responsibly reporting this issue. After careful investigation, we determined that the access patterns demonstrated are consistent with intended protections and existing controls, and do not represent a bypass of a security boundary or unauthorized access to data," a Microsoft spokesperson told Ars. "The authorization period has a timeout and anti-hammering protection that limit the impact of malicious queries."

OpenAI is updating Codex with more agent-like capabilities, positioning it as a more direct rival to Anthropic's Claude Code. Some of the new features include the ability to operate macOS desktop apps, browse the web inside the app, generate images, use new workplace plug-ins, and remember useful context from past tasks. The Verge reports: Codex will now be able to operate desktop apps on your computer, OpenAI says in a blog post announcing the update. It can work in the background, meaning it won't interfere with your own work in other apps, and multiple agents can work in parallel. For developers, OpenAI says "this is helpful for testing and iterating on frontend changes, testing apps, or working in apps that don't expose an API." The feature will start rolling out to Codex desktop app users signed in with ChatGPT today and will initially be limited to macOS. OpenAI did not indicate a timeline for when use will expand to other operating systems. EU users will also have to wait, it said, adding that the update will roll out to users there "soon."

Codex is also getting the ability to generate and iterate on images with gpt-image-1.5, new plug-ins for tools like GitLab, Atlassian Rovo, and Microsoft Suite, and native web browsing through an in-app browser, "where you can comment directly on pages to provide precise instructions to the agent." OpenAI also said it will also be easier to automate tasks, with users able to re-use existing conversation threads and Codex now able to schedule future work for itself and wake up automatically to continue on a long-term task. Codex will also be getting a memory feature allowing it to remember useful context from past experience, such as personal preferences, corrections, and information that took time to gather. OpenAI said it hopes the opt-in feature, which will be released as a preview, will help future tasks complete faster and to a quality that previously required detailed custom instructions. The personalization features will roll out to Enterprise, Edu, and EU users "soon."

An anonymous reader quotes a report from Reuters: Alphabet's Google is negotiating an agreement with the Department of Defense that would allow the Pentagon to deploy its Gemini AI models in classified settings, the Information reported on Thursday, citing two people with direct knowledge of the discussions. The two parties are discussing an agreement that would allow the Pentagon to use Google's AI for all lawful uses, according to the report.

During the negotiations, Google has proposed additional language in its contract with the department to prevent its AI from being used for domestic mass surveillance or autonomous weapons without appropriate human control, the Information reported. The Pentagon will continue to deploy frontier AI capabilities through strong industry partnerships across all classification levels, a Pentagon official said, without confirming any talks with Google.

Anthropic released Claude Opus 4.7, calling it its strongest generally available model and an improvement over Opus 4.6 in areas like software engineering, instruction-following, tool use, and agentic coding. But the company says it is "less broadly capable" than the restricted Claude Mythos Preview, "which Anthropic rolled out to a select group of companies as part of a new cybersecurity initiative called Project Glasswing earlier this month," reports CNBC. From the report: The launch of Claude Opus 4.7 on Thursday comes after Anthropic launched Claude Opus 4.6 in February. Anthropic said the new model outperforms Claude Opus 4.6 across many use cases, including industry benchmarks for agentic coding, multidisciplinary reasoning, scaled tool use and agentic computer use, according to a release. Anthropic said it experimented with efforts to "differentially reduce" Claude Opus 4.7's cyber capabilities during training.

The company encouraged security professionals who are interested in using the model for "legitimate cybersecurity purposes" to apply through a formal verification program. Claude Opus 4.7 is available across all of Anthropic's Claude products, its application programming interface and through cloud providers Microsoft, Google and Amazon. The new model is the same price as Claude Opus 4.6, Anthropic said.

An anonymous reader quotes a report from UploadVR: A group of independent researchers built a device that can artificially induce smell using ultrasound, with no consumable cartridges required. [...] The team of four are Lev Chizhov, Albert Yan-Huang, Thomas Ribeiro, Aayush Gupta. Chizhov is a neurotech entrepreneur with a background in math and physics, Yan-Huang is a researcher at Caltech with a background in computation and neural systems, and Ribeiro and Gupta are co-researchers on the project with software engineering and AI expertise.

Instead of targeting your nose at all, the device directly targets the olfactory bulb in your brain with "focused ultrasound through the skull." The researchers say that as far as they're aware, no one has ever done this before, even in animals. A challenge in targeting the olfactory bulb is that it's buried behind the top of your nose, and your nose doesn't provide a flat surface for an emitter. Ultrasound also doesn't travel well through air. The solution the researchers came up with was to place the emitter on your forehead instead, with a "solid, jello-like pad for stability and general comfort," and the ultrasound directed downward towards the olfactory bulb.

To determine the best placement, they say they used an MRI of one of their skulls to "roughly determine where the transducer would point and how the focal region (where ultrasound waves actually concentrate) aligned with the olfactory bulb (the target for stimulation)". [...] According to the researchers, they were able to induce the sensation of fresh air "with a lot of oxygen", the smell of garbage "like few-day-old fruit peels," an ozone-like sensation "like you're next to an air ionizer," and a campfire smell of burning wood. While technically head-mounted, the current device does require being held up with two hands. But as with all such prototypes, it likely could be significantly miniaturized.