DeepSeek’s V4 Flash has topped model leaderboards and been hailed by developers as a “total monster” since its rollout. But in real-world testing, it completed just 53.8% of a batch of complex agent tasks.
Composio ran the model through eight different agent harnesses, including Claude Code, Codex, and OpenCode, on 30 deliberately difficult, multi-step tasks spanning live tools like Gmail, GitHub, Slack, and Google Sheets. Of 240 total runs, 129 passed — and only six of the 30 workflows were completed successfully by every harness tested.
The gap illustrates why orchestration, not raw model capability, may decide whether the model succeeds in enterprise settings: the same model produced substantially different results depending on the harness, tool configuration, caching behavior, retries, and provider stack it ran on.
DeepSeek said it will be hiking the prices for V4 Flash and Pro, models that have quickly become favorites among developers building coding assistants and agents.
While it seems the move might undercut its very appeal — strikingly capable models at ultra-low pricing that frontier providers simply can’t match — it also moves the story beyond the now-clichéd “cheap Chinese model” narrative, as early use cases emerge and enterprises figure out where different models fit into their tech stacks and what workflows they should be aimed at.
DeepSeek rolled out V4 Flash to public beta July 31, and made V4 Pro generally available on August 13. The 284-billion-parameter Flash is built for volume and speed, the 1.6 trillion-parameter Pro for more complex workflows.
Both models have flexible reasoning capabilities (low, high, max) and ‘thinking modes’ applying chain-of-thought (CoT) reasoning to improve answer accuracy.
Users were immediately impressed by Flash’s capabilities. It has dominated OpenRouter’s usage leaderboard since its rollout, currently the most-used model on the platform by weekly token volume.
“The adoption numbers of the initial DeepSeek V4 Flash were insane,” ML researcher Nathan Lambert posted to X, adding that the new version “scored the same as GLM 5.2,” making it a “total monster” that will be used extensively.
DeepSeek switching the cost model adds an interesting dimension.
V4 API rates are going up by as much as 1,100% depending on the model, token type and time of use. The new pricing structure:
Flash will be 22 cents per million input tokens and 66 cents per million output tokens off-peak; and 44 cents per million input tokens and $1.32 per million output tokens at peak. This represents a 57% to 371% increase.
Pro will be 66 cents per million input tokens and $1.98 per million output tokens off-peak; and $1.32 per million input tokens and $3.96 per million output tokens at peak. This shows a 51% to 355% jump.
Cache hits, meanwhile (when models reuse prompts rather than starting from scratch), are going up between 52% and 1,100%.
DeepSeek says offering 50% lower off-peak usage is intended to encourage “more flexible workload scheduling.” Seventeen of every 24 hours stay at half price, and the new structure actually prices the company’s home market the highest.
“This is not a simple price rise,” said Sanchit vir Gogia of Greyhound Research. “It is a pricing architecture that makes the timing of inference an economic variable.”
Work that can wait — such as batch evaluation, synthetic-data generation, and overnight development runs — moves into the cheap hours; interactive agents and live operations cannot. Gogia said irritation among developers and enterprises is genuine and vocal, and that DeepSeek’s past low pricing doesn’t obligate it to stay cheap forever.
At first glance, it does look like a “suicidal move from a platform still looking for credibility against more established AI model vendors,” said tech analyst Carmi Levy. The increases will certainly eat into DeepSeek’s price advantage and force customers to weigh concerns around the company’s Chinese origins more heavily.
Still, DeepSeek remains far cheaper by all pricing measures relative to competing models from OpenAI, Anthropic, Google, Cohere, xAI, and others, he said.
So, while the move will force DeepSeek to emphasize performance and security over cost, it hardly wipes out its already-notable price-performance advantage, and still gives customers ample wiggle room to justify its use for specific workloads, Levy said. The math will just have to be more tightly calculated.
“The advantage will likely erode over time as DeepSeek inevitably continues to align pricing with market realities, but for now it’s still easy to make the business case,” Levy said.
Adoption inside enterprises remains an open question due to cost, capability, reliability, data governance, security, and other factors.
One use case is batch processing, Levy said. This kind of work is typically routine and repetitive rather than demanding top-tier intelligence, so it makes sense to use a cheaper, more efficient model. “It’s a high-performance inference engine that enterprises can consider using for point solution workloads rather than as a wholesale replacement for the incumbent offerings,” Levy said.
Partial adoption will likely involve isolated, non-sensitive workloads with clearly-defined success metrics, strict oversight, permissions controls, and fallback models in case of failures, he noted. Broader deployment will require DeepSeek and its hosting partners to demonstrate strong reliability, security, privacy, auditability, and deployment options. As it adjusts price structures based on demand, DeepSeek also must retain a large enough price-performance advantage to justify any risk, he said.
Expect unsanctioned, smaller-scale use in backroom labs and contained test environments as IT teams get familiar with the new model and figure out when and how to bring it to senior leadership for budget approval.
“DeepSeek has built a well-earned reputation as a global disruptor,” he said, “and it’s clear that its march to broader enterprise adoption will continue to gather momentum.”
While many use cases are still in the experimental phases, Meta software engineer Naman Ahuja offers one that could translate directly into enterprise environments. In a project unrelated to his employer, he built a home-automation agent with DeepSeek V4 Flash to explore how a lower-cost model performs as the reasoning/orchestration layer for a real multi-tool workflow.
When he leaves home, an agent coordinates several actions across otherwise separate systems: Such as setting a thermostat to “away” to reduce unnecessary energy use, arming a Ring security system, closing and locking doors.
“What interested me was not simply whether the model could understand a command, but whether it could translate intent into a sequence of actions across multiple tools where reliability matters,” Ahuja said.
The biggest lesson was that once a model can take actions, reliability matters as much as intelligence. The system needs structured tool outputs, verification that actions actually succeeded, retry/failure handling, and clear boundaries around what the model is allowed to do.
In the case of enterprise, “the architecture is similar.” Home devices change to ticketing systems, databases, CRM platforms, or infrastructure APIs. The most useful agents will likely orchestrate repetitive workflows across multiple systems, with scoped permissions, auditability, observability, and human approval for higher-risk actions.
“Many valuable AI agents will not be chatbots; they will be background agents coordinating APIs, infrastructure, and business systems in response to events,” he said.
But Flash’s API is still in public beta, Gogia pointed out, and there is not yet an evidence trail of settled enterprise adoption, real-world deployments, and named customers.
“The benchmark story is looser than its retelling, the portfolio story is newer than it looks, and the economics have moved into the system around the model,” he said. Developer mainstreaming is proven; enterprise standardization is not. “The model is mainstream by traffic and still unproven by contract.”
DeepSeek’s own integration guidance is an important consideration, he said: Its documentation for at least one popular agent environment states that built-in V4 entries are not sufficient for reliable operation without compatibility overrides.
“Which is a vendor telling the market, accurately, that benchmark performance is not a proxy for production readiness,” Gogia said. “A model can score beautifully and still misbehave once tools, credentials, and state enter the room.”
The serving layer behaves no differently: the same open weights run by different hosts show visible differences in throughput and uptime. “Choosing Flash therefore answers one procurement question and opens three more: Who serves it, where it runs, and which controls surround it,” Gogia said.
DeepSeek offers a nuanced case for a multi-model future. V4 Flash is being deployed as the high-volume worker inside diverse estates, Gogia noted: It handles routine generation, retrieval, and background automation, while more difficult or sensitive tasks go elsewhere.
“The question is whether its performance is sufficient for the real-world workflows enterprises actually run, not whether it tops every benchmark,” he said. Enterprises must determine which combination of model, harness, and provider completes the work safely at the lowest cost.
Adam Dalloul, CEO and founder of EmpirioLabs AI, pointed out that bigger isn’t always better; workflows should be task-dependent. For example, his team at EmpirioLabs AI — which hosts 100-plus models on one API, including DeepSeek V4 Flash — were recently working on translating its site into different languages, and there was no need for a large model like GPT 5.6 Sol or Opus 5 to complete the task.
“This is where subagents come in handy,” he said.
His recommended approach: Spawn cheaper subagents and adapt per task. For instance, use Flash variants for day-to-day work, and Pro variants when you need something more powerful. “It depends on the nature of your application.”
Many companies are pivoting towards their own internal benchmarks to route models effectively, Dalloul noted. For example, his team has a workflow that puts a model through various gates and instructions. This helps them identify the model with the speed and accuracy required for the task.
In another example, one of his enterprise clients exclusively wanted access to DeepSeek V4 Flash. They had tested a variety of models and V4 Flash was the only one that met their criteria for speed, cost, and an “appropriate intelligence threshold.”
Meta’s Ahuja agreed that smaller, more efficient models can handle frequent, well-defined agentic tasks, while more expensive frontier models can be reserved for “ambiguous, difficult, or higher-risk decisions.” The relevant metric increasingly becomes cost per successfully completed workflow rather than simply cost per token.
The trade-off, however, is that cheap inference does not automatically mean cheap or safe automation, he said. Once an AI system can take actions, reliability, verification, permissions, failure handling, and security become much more important.
“A failed text response is inconvenient; a failed action in an operational workflow can have real consequences,” Ahuja said.
Alibaba says its Qwen models have passed 3 billion downloads, making them the most used open-weight family in the world. The Hugging Face report published a day earlier counts 2,045 million for Qwen this year, and 151,448 derivatives against Alibaba’s claim of more than 300,000.
Alibaba says its Qwen models have passed 3 billion downloads, putting the family ahead of anything from Meta or Google. It is the clearest claim yet that Chinese labs are winning the open-weight race.
The report it rests on says something slightly different. Hugging Face’s State of Open Models, published the day before, counts 2,045 million Qwen downloads this year, or 2,061 million including every repository. The 3 billion figure came from Alibaba’s own emailed statement.
The gap does not change who is first. It does mean the headline number is about a third larger than the platform being cited actually recorded.
The same pattern appears in the derivatives. Alibaba says its ecosystem has produced more than 300,000 derivative models, while Hugging Face counts 151,448 Qwen-based derivatives on its Hub, which is still 2.6 times Meta’s entire footprint there.
This is familiar ground. TNW noted earlier that Alibaba called Qwen3.8 the world’s second-best model and showed no proof.
Hugging Face is careful about what its numbers mean. Downloads measure activity inside its own hub and do not capture API usage, private deployments or models distributed elsewhere, and Alibaba runs its own distribution platform in ModelScope.
None of which touches the direction of travel. “Qwen has become part of the default workflow for developers deciding what models to fine-tune and deploy,” the report says.
The method is volume and reach. Alibaba has open-sourced more than 460 models under permissive licences and pushes them through its cloud into Southeast Asia and Africa, though it has also said it wants to charge the heaviest users of those open models.
American labs are answering in kind. Meta and Nvidia have both shipped new open models in recent weeks, and Nvidia has been explicit that it is chasing China on open weights.
The bigger threat to Qwen’s run may be domestic. Beijing has been weighing curbs on overseas access to the country’s best models, which would take the reach that produced these numbers away.
A new information-stealing malware called AmnesiaStealer, which targets macOS users via ClickFix attacks, includes a streaming module that allows the attacker to interactively control the victim’s web browser.
A notable capability is copying the victim’s Chromium profile, including its authentication state, and loading it into a hidden, headless browser on the infected system.
This allows the hacker to access victims’ authenticated sessions while preserving the identifiers associated with the browser, host, and network.
AmnesiaStealer can collect data in 16 Chromium-based web browsers as well as other sensitive information, such as passwords, cryptocurrency wallets, Apple Notes and documents, and keychain data.
The malware is currently distributed through ClickFix campaigns that use a fake GitHub download page to drop a password-protected ZIP archive.
Researchers at Jamf, an Apple device management and security company, analyzed AmnesiaStealer’s distribution and found that it used the same template previously used to spread the Atomic and MacSync infostealers.
The ClickFix command executes a shell-script loader that downloads and launches the password-protected archive containing the AmnesiaStealer Mach-O payload.
The malware captures the victim’s macOS password and uses it to collect keychain data, as well as browser profiles, Apple Notes, Telegram sessions, documents, system information, and cryptocurrency wallet data.
The researchers highlight that the malware features a component called stream_module, retrieved using the remote_stream command, which gives the malicious operator remote control over authenticated sessions deployed from a headless browser instance.
According to Jamf, AmnesiaStealer’s stream_module can duplicate user profiles in seven Chromium-based browsers, including Google Chrome, Microsoft Edge, Vivaldi, Arc, Opera, Brave, and Chromium, because they share the same DevTools Protocol, launch flags, and cookie encryption.
The module launches the legitimate browser executable in headless mode with command-line switches that weaken browser defenses, duplicates the victim’s profile, and specifies its location for storing the profile data.
The malware then establishes a WebSocket channel that connects to the operator’s relay and sends a JSON registration message containing the browser name and build.
The operator can then send commands over this channel, such as navigation and mouse clicks, while the malware returns status and tab information as JSON and transmits screencast frames as binary WebSocket messages.
A second WebSocket channel connects to the local headless Chromium instance through the browser’s webSocketDebuggerUrl, providing access to the Chrome DevTools Protocol (CDP).
This allows the hacker to navigate websites with mouse and keyboard control, export or import cookies, and operate online portals using the victim’s existing authenticated sessions.
“The operator receives a live screencast of the session at around 3fps and can drive it with a full input set: keyboard, mouse, scroll, navigation and tab management,” Jamf explains.
“In effect the remote_stream command turns an infected host into a live, operator-driven browser running the victim’s authenticated sessions, which is a materially different level of access from file collection.”
According to the researchers, the AmnesiaStealer can exfiltrate cookies, saved logins, browsing history, bookmarks, extensions, local state, and other profile data from the 16 Chromium-based browsers it targets.
It also steals cryptocurrency wallet details and identifies them by enumerating extensions and IndexedDB data.
Jamf notes that the malware contains a fallback mechanism when it runs on macOS 26 and cannot recover the existing Chrome Safe Storage key, which replaced it with an attacker-supplied value.
This makes previously stored cookies and passwords permanently unreadable while allowing the attacker to decrypt data later.
The Chrome DevTools Protocol (CDP) has been abused by malware in the past, including by Chaos ransomware to hide command-and-control communications, and by Chaes malware to expose browser functions that could enable data theft.
However, AmnesiaStealer appears to be the first documented macOS malware to combine a cloned Chromium profile with CDP-based, live remote control, allowing attackers to interact with authenticated sessions through a hidden browser running on the infected computer.
Users are advised never to execute commands in the terminal that they found online and don’t fully understand.
Overall prevention scores can hide what happens after initial access. Once attackers are using valid credentials, prevention drops sharply.
The Blue Report 2026 measures defenses technique by technique across 338 million simulations run in customer production environments.
We may receive a commission on purchases made from links.
Anker has become synonymous with phone chargers and power banks. It’s one of the most reputable brands and is quite popular in global markets like North America and Europe. The company is headquartered in China and has over 30 offices that operate worldwide. Anker got its start selling replacement laptop batteries on Amazon but quickly expanded into smartphone chargers, cables, and other charging accessories.
A quick search on Amazon for a “power bank” or a “phone charger” is highly likely to yield at least a few of Anker’s products among the top results. The reason Anker is so popular comes down to a combination of quality, consistency, and, most importantly, value. Apple charges $60 for its 35W dual-port USB-C adapter, while Anker sells a 65W 3-port charger for half the price.
The company is currently best known for its high-capacity power banks, such as the Anker 25,000 mAh Power Bank, or its fast charging wall adapters, like the Anker 140W GaN Charging Block. Both of these products have high ratings and thousands of reviews on Amazon, but you’d be surprised to know that Anker’s catalog doesn’t stop at charging accessories for laptops and smartphones. Over the years, the brand has expanded into categories such as smart home accessories, home entertainment, wireless audio, and portable power solutions for outdoor use. Here’s a closer look at a few brands you may not have known are owned by Anker.
Anker has already solved many of the problems consumers faced with expensive, bulky phone chargers and power banks. The next logical step was addressing the battery problem at a much larger scale — home energy. In 2023, the company decided to apply its expertise in power storage and delivery to home energy by launching Anker Solix — a brand dedicated to portable power stations and solar generators.
While some of Anker’s larger power banks can keep a laptop going for a couple of hours, Solix offers solutions for bigger appliances that require a lot more power. Think refrigerators, televisions, and power tools — essentials that are often powered directly from a wall outlet. At this scale, batteries are measured in watt-hours (Wh) and aren’t touted for how many times over they can charge your phone but for how long they can keep larger appliances running.
The Anker 536 PowerHouse is a 508Wh portable power station that uses LiFePO4 (lithium iron phosphate), a composition that offers a higher power-to-weight ratio than traditional lead-acid batteries. Anker Solix also sells home battery systems that are designed to keep your entire house powered during an outage. If you ever decide to install solar panels, Solix has a few solutions for that, too. Despite operating in a more specialized category that requires larger-scale power storage solutions, Anker Solix’s products remain competitively priced. Most of its portable power stations come with a couple of AC outlets, USB ports, and surprisingly fast charging for when they run empty.
Hybrid work setups have grown in popularity ever since the pandemic hit. This helped drive the comeback of better-quality webcams and headsets with microphones that don’t sound like you’re underwater. AnkerWork is the company’s answer to this growing demand. The sub-brand’s portfolio of products is currently not as rich as some of Anker’s other ventures, but there’s not much to complain about here.
The PowerConf C200 Webcam brings a noticeable improvement to your video calls with its 2K resolution and wider field of view. It snaps on top of your monitor, comes with two microphones, and has an adaptive light sensor that adjusts how your image looks based on the ambient light around you. For the security-conscious, the webcam offers a switch that physically covers the camera lens for when you don’t need it. If you don’t like how high your monitor is or simply prefer a different angle, you can grab AnkerWork’s Mini Tripod that can hold a webcam, smartphone, or a dedicated camera.
The company also sells speakerphones that are designed to dramatically improve how you sound to others in meetings, as well as how well you can hear them. The AnkerWork SR500 uses eight microphones to handle audio recording from all directions and a 10W speaker that fills the room with sound. You can connect up to five of these speakerphones if you’re in a larger meeting room with several people.
Eufy may be a name you recognize. It specializes in smart home surveillance and cleaning, with a wide range of security cameras, video doorbells, and robot vacuums to choose from. Anker launched the sub-brand in 2016 as part of its efforts to expand beyond the charging accessories it was known for. Looks like it worked — Eufy’s products generally enjoy positive reviews despite stiff competition in the home automation space.
What Eufy is doing a bit differently in this segment is reducing the reliance on subscriptions. Popular competitors like Ring and TP-Link often require you to subscribe to a monthly plan if you wish to access longer video histories or unlock more advanced features. While this makes the upfront purchase more affordable, the ongoing costs can add up over time. By contrast, the Eufy Indoor Camera that’s priced at $70 doesn’t ask for any monthly fees and instead lets you store your footage locally via its SD card slot.
That said, Eufy’s security controversy in 2022 raised concerns about the company’s privacy claims. The company responded by clarifying how certain features and data flows worked. Eufy continues to make smart home devices, with some of its newer security cameras offering features like motion detection and the option to store your recordings in the cloud. Some of its outdoor security cameras are solar-powered, too. Eufy also ranks quite well among major robot vacuum manufacturers.
Like Eufy, Soundcore is one of Anker’s more recognizable brands. It was officially unveiled as an audio-focused sub-brand in 2018, though Anker had been selling Bluetooth speakers and other audio products for years before that. Today, Soundcore has a decently large catalog of over-ear headphones, wireless earbuds, speakers, and even an AI voice recorder for some reason. Though it’s Soundcore’s affordable range of headphones and earbuds that gets the most attention, the brand does have a couple of premium offerings as well.
The Soundcore Space One Pro headphones retail for $200 and feature support for Hi-Res audio, active noise cancellation, Bluetooth multipoint connectivity, and up to 60 hours of playback time with ANC off. Users generally say good things about Soundcore as a brand, too, with its products punching well above their price range. That said, if you’re looking for a true premium alternative to headphones like the Sony WH-1000XM6 or Apple’s AirPods Max, Soundcore doesn’t operate in quite the same league yet.
You do get an assortment of earbuds here. At $230, the Liberty 5 Pro Max are Soundcore’s flagship true wireless earbuds, featuring an AMOLED display on the case that lets you manage earbud settings in case you loathe having to pull out your phone every time. Soundcore also sells earbuds specially designed for sports and sleep. You have a wide range of open-ear earbuds to choose from, too. Its selection of Bluetooth speakers includes options that are waterproof, portable, or decked out with features like spatial audio and Hi-Res sound.
Having locked down charging accessories, smart home products, and consumer audio, Anker turned to the world of home entertainment with Nebula — a sub-brand focused primarily on delivering portable projectors. Nebula has been making projectors since 2017, but the sub-brand joined Soundcore in 2025. Like some of its other sub-brands, Soundcore Nebula has options at various price points.
The Nebula P1i is the entry point to the lineup, allowing users to project movies, TV shows, and other presentations in up to 1080p, all while being relatively portable. It comes with two 10W speakers that can be flipped around for a more immersive audio experience. You get essentials like auto Keystone correction, a tiny adjustable stand, and built-in support for Google TV so you don’t need to plug in an input source.
Moving up the lineup, you’ll find larger, more capable projectors that can serve as replacements for televisions. Products in Nebula’s Mars and Cosmos series offer resolutions up to 4K, much larger projection sizes, and technologies like Dolby Vision. Besides projectors, Nebula also sells accessories like tripods, carry cases, and projector screens.
Nate Scovill decided the Red Bull slogan needed a proper test. He bought cans of the stuff, cleared space in his workshop, and set out to see whether an energy drink could actually launch something into the air. His recent video walks through three attempts, each more serious than the last, until one of them finally left the ground under its own power.
He started simple, with a can of Red Bull, which included 160 calories and 38 grams of sugar. After drinking it, he attached some cardboard fins to the empty can and tossed it into the air like a smart paper plane. Not surprisingly, the can flew barely a few feet at most. He added up the calories from drinking it as energy consumed and declared the entire thing technically valid, even if it seemed a little like cheating. It made for a humorous start and set the tone for the rest of the project.
The next step was a proper-looking rocket. He took old soda bottles, cut and taped them into a pressure vessel, tacked on some cardboard fins, painted them red, blue, and silver, and created a primitive valve out of a wine cork, heat-shrink tubing, and a soccer ball needle. Nate filled the bottle with liquid, inflated it with a bike pump, and launched it into the woods. It went well until one of the fins snapped off. Nonetheless, the propulsion system was functioning perfectly, with compressed air driving the liquid out the nozzle. In the end, the Red Bull only added weight and color, which wasn’t too bad. So, in spirit, if not in reality, this was a pretty impressive water rocket powered by an energy drink.
The third attempt was when the magic happened. Old-school amateur rocketeers swear by rocket candy, which uses sugar as fuel and potassium nitrate (often sold as stump remover) as oxygen. When properly blended and put into a tube, the mixture produces a tremendous thrust. Scovill reasoned that since the Red Bull already had sugar dissolved in it, it would suffice if he could only get all the other things out of the way.
So he cooked it down in a pot at low heat until most of the water evaporated, leaving a thick syrup. Water is a major impediment to rocket fuel performance, hence this was a necessary step. However, there was still one issue to resolve, since the carbon dioxide remained trapped in the liquid. Bubbles in the mix are a real pain when you’re attempting to get a nice burn going. He solved the problem by putting the reduced liquid in a vacuum chamber and sucking out the gas. Then, using a coffee grinder, he finely ground the potassium nitrate and blended it with the syrup until it was like putty. The final combination was put into little tubes he’d created from rolled up newsprint. To provide some further insurance, he employed a cat litter base as both a safety buffer and an igniting assist. Finally, he created a small control panel with a key switch and a large red button, connected it to a model rocket igniter, and tested three different batches.
The first bunch of rockets refused to move, while the second one caught fire, but the paper tube it was in was so badly wrinkled that the pressure could escape sideways. As a result, flames erupted from the side, melting the outer nozzle, and the rocket sat still on the pad. On the third attempt, this time with a fully degassed syrup, they had it right. When Scovill clicked the ignition button, the motor screamed to life, the rocket jumped off the rail, and for a few seconds, a homemade motor powered by a Red Bull energy drink flew under its own power. To commemorate the occasion, he even created a mission patch.
When the water and gas are removed, the chemistry becomes much more straightforward. When the sugar molecules break down under heat and react with the oxygen released by the nitrate, they form an expanding hot gas that propels the rocket onward. It’s as simple as that, but the true issue is getting the ratio precisely right and keeping the moisture under control. That’s where Red Bull comes in; it’s not 100% sucrose, so determining the proper proportions requires some trial and error. Scovill’s successful motor demonstrated that the overall concept could function even with all of the additional additives contained in a commercial energy drink.
[Source]
When Soundgarden Wrote “Black Hole Sun,” they may have inadvertently been onto something.
In a new paper in Nature, scientists have chronicled a “black hole star” that’s the size of our solar system and glows bright red in the depth of space. The findings are based on observations from the James Webb Space Telescope, which continues to change how astronomers think about the cosmos.
Researchers focused on some of the most distant observable phenomena in the universe, which appear to be something like a star in disguise. Based on the light they emit, they appear to share some characteristics with giant stars, but astronomers suspect they are actually black holes shrouded in an extraordinarily dense and luminous gas.
These hypothetical objects have been dubbed black hole stars, and the new research, published Wednesday, lays out the case for them.
The signal that’s at the center of the study comes from a time when the universe was just 660 million years old. (It’s now almost 14 billion years old.) In the JWST images, the object appears as one of the many bright, reddish spots that the telescope has found in the early universe.
From a distance of billions of light-years, this phenomenon appears to be a gigantic star. But the object, known as MoM-BH*-1, has a few curious properties. It emits about 100 billion times more energy than any star could produce through nuclear fusion, according to the team. Even for the era in which it formed—during the early universe—it is an extremely bright object.
Furthermore, part of its light abruptly disappears. Stars can produce a similar signal when their atmospheres absorb certain wavelengths, but in MoM-BH*-1, the effect is too intense to be explained as normal star behavior. Researchers believe that an enormous amount of extremely dense gas is absorbing that light.
These clues led scientists to the black hole star hypothesis. At the center of MoM-BH*-1, there would be a black hole devouring matter and releasing enormous amounts of energy. Surrounding it, a massive gas envelope. The energy would pass through that gas before escaping into space and, in the process, would acquire some of the characteristics we would normally expect to find in starlight.
The hypothesis could help address another mystery uncovered by the JWST. The telescope finds, with surprising frequency, small red spots in the early universe whose origins remain uncertain. They don’t fit neatly into the usual categories of objects in deep space. Scientists have put forward a number of explanations, including that they’re compact galaxies teeming with stars, black holes hidden by dust, and systems completely different from those we observe in the present-day universe.
Oil giant Shell has confirmed it is investigating a potential security incident after the Clop ransomware gang claimed it stole 89GB of data.
Shell is a British multinational energy conglomerate and one of the world’s top three oil and gas companies, after Chevron and ExxonMobil. It has 85,000 employees in more than 70 countries and operates a massive network of tens of thousands of service and recharge stations that serve over 20 million customers daily.
According to a recent post on Clop’s dark web data leak site, the allegedly stolen files include engineering drawings, scans of facility testing reports, photos of the facilities, and project plans.
“We are aware of a potential incident. We are working with our security teams and relevant experts to investigate,” a Shell spokesperson told BleepingComputer when asked to confirm Clop’s data theft claims.
While the company has yet to share more information, the Clop gang listed it on its leak site as one of 43 new victims likely targeted in data theft attacks against Internet-exposed PTC Windchill and FlexPLM instances exploiting a critical improper input validation vulnerability tracked as CVE-2026-12569.
As part of the same attacks, Clop also claimed it stole sensitive data, including backups, system files, projects, drawings, diagrams, and blueprints, from the networks of tech conglomerates General Electric and Philips.
GE and Philips spokespersons were not immediately available for comment when BleepingComputer contacted them earlier today. A PTC spokesperson has also yet to reply to a request for comment.
PTC began releasing CVE-2026-12569 security patches on June 17 and, even though it didn’t confirm in-the-wild exploitation, it also released a private advisory urging customers to review environments for indicators of compromise (IOCs).
After PTC warned customers of “heightened threat activity” on June 26, the U.S. Cybersecurity and Infrastructure Security Agency (CISA) also confirmed that the flaw is actively exploited in attacks, adding it to its Known Exploited Vulnerabilities catalog, and ordering federal agencies to secure their PTC Windchill and FlexPLM instances within three days.
CVE-2026-12569 also prompted emergency action from German authorities, with the Federal Office for Information Security (BSI) warning PTC customers in the middle of the night to patch their systems as quickly as possible.
Clop’s Windchill and FlexPLM attacks were also confirmed by the Ransomware Information Sharing and Analysis Centre (Ransom-ISAC), a non-profit organization dedicated to the tracking and defense against ransomware threats, and by cybersecurity company ReliaQuest, which said that the threat actors have been deploying JSP webshells that allow them to steal sensitive data from victims’ compromised PLM platforms.
PTC FlexPLM and PTC Windchill are enterprise software platforms in the Product Lifecycle Management (PLM) category, used to track, design, and manage products up to final manufacturing.
The two systems are widely popular among engineering, manufacturing, quality, and supply chain teams at high-profile companies across the aerospace, defense, automotive, heavy machinery, retail, and medtech sectors. PTC says that its products are used by over 30,000 customers globally, including more than 1,500 brand and retail customers using FlexPLM.
Overall prevention scores can hide what happens after initial access. Once attackers are using valid credentials, prevention drops sharply.
The Blue Report 2026 measures defenses technique by technique across 338 million simulations run in customer production environments.
Since the end of last month, several water utilities in the United States have been hit by cyberattacks, causing alarm in the country.
For years, water utilities and other critical infrastructure facilities in the power sector, for example, have been targeted by hackers, whether government-backed or individuals. What makes these recent attacks — allegedly carried out by Iran — particularly concerning is how widespread they were, hitting targets in around a dozen states.
The U.S. has more than 150,000 water systems, some of them run by local companies. In theory, that should make it harder for hackers to target several facilities at the same time. But on the flip side, the companies running these systems may not have the resources or cybersecurity expertise needed to protect themselves.
Cybersecurity experts have long believed that Iranian hackers target low-hanging fruit in opportunistic isolated attacks, so this hacking campaign could be a significant escalation.
A lot has happened since news of the initial attacks broke two weeks ago. So we decided it was a good time to recap what we know so far, and what we don’t.
On July 28, Minnesota authorities announced that water treatment plants in more than 30 communities were hit by coordinated cyberattacks.
Two days later, the FBI said water and wastewater utility companies in “at least seven states” reported incidents, and in some cases the attacks “degraded water operations.” Since then, apart from Minnesota, there have been reported hacks against water facilities in Arkansas, Georgia, New Jersey, and Michigan.
The short answer is: we don’t know yet, but the No. 1 suspect is the Iranian government.
As of today, officially, the U.S. government has yet to name the culprit behind the coordinated wave of hacks.
However, the first incidents in Minnesota came days after the U.S. Cybersecurity and Infrastructure Security Agency (CISA) warned that Iranian hackers were targeting internet-connected devices in water systems and the energy sector, without saying where those attacks were occurring. (CISA had originally published this warning in April, and updated it before the Minnesota attacks.)
After the initial wave was uncovered in Minnesota, President Donald Trump said he did not think “there was an Iranian cyberattack.” Instead, he blamed the state, perhaps because it is run by democratic governor Tim Walz, who was chosen as Kamala Harris’ vice president candidate in the 2024 elections.
Trump’s claim came a day after Wired reported that the Water Information Sharing and Analysis Center, or WaterISAC, a nonprofit group that distributes cybersecurity information among the water sector, told its members that the recent attacks “aligned” with the hacking campaign CISA warned of — effectively accusing the Iranian government.
Earlier this week, The Washington Post reported that U.S. intelligence agencies “are confident” that Iran, and in particular the Islamic Revolutionary Guard Corps (IRGC), is responsible. The attribution isn’t public yet, according to the paper’s sources, because the agencies are not sure which specific unit within the IRGC was responsible, and also because officials may be reluctant to contradict Trump’s claim.
Iranian government hackers have a history of targeting critical infrastructure in the U.S., and it’s possible that these attacks are part of its strategy to retaliate against the country because of the six-month war.
Until now, Iranian hackers had only limited success in their cyberattacks against U.S. targets. In March, a hacktivist group called Handala disrupted the operations of medical tech giant Stryker. The U.S. government later accused Handala of being operated by Iran’s Ministry of Intelligence and Security (MOIS). Then, the group claimed responsibility for hacking the personal Gmail account of FBI director Kash Patel.
The reality is that some systems inside critical infrastructure facilities are exposed to the internet and relatively easy to find. Earlier this month, cybersecurity firm Forescout reported finding more than 2,800 controllers in U.S. water systems exposed online. If a system is exposed, it doesn’t automatically mean hackers can take over control and cause real-world effects. But that has happened in some isolated cases in recent attacks.
The FBI said some of the cyberattacks around the country caused loss of pressure, which “could potentially allow untreated groundwater to seep into pipes,” and flooding.
The town of Braham in Minnesota, one of the first to report an incident, had to take its water plant offline for a few hours, urging its around 1,700 residents to conserve water. The city of Maple Plain, also in Minnesota, briefly declared a state of emergency. In a county outside Atlanta, Georgia, local officials briefly told residents to boil water before using it as a precautionary measure.
The worst effect, however, may be psychological. These attacks have been widely covered in national and local press, causing people to worry about the safety of a fundamental and basic need like water. That may very well be part of the hackers’ goals: to spread panic and fear.
When you purchase through links in our articles, we may earn a small commission. This doesn’t affect our editorial independence.
Google’s Pixel collection has shifted from camera- to AI-focused in the past few years, driven mainly by the ever-more-capable Gemini – and that, of course, continues with this year’s Pixel 11 collection.
However, if you’re tempted to go out and pre-order one ahead of release later this month, I’d advise you to think twice.
Why? Because some of the touted AI features that Google has been singing and dancing about since the initial announcement of Android 17 back in May – arguably the reason to upgrade from last year’s Pixel 10 – aren’t actually available at launch in regions outside of the US.
The Pixel 11’s AI features fall under the new ‘Gemini Intelligence’ umbrella that Google first outlined at the announcement of Android 17 back in May – and there are plenty on the way.
It starts with big upgrades to existing features like Rambler, which works much like the current speech-to-text on Android, but with the ability to remove filler words, make edits on the fly and more, which should make it way easier to get your point across using your voice without looking stilted or awkward.
There’s also a new AI-infused lock screen experience, with Google explaining that it’ll be able to, say, proactively display a restaurant’s name that you’re going to, along with AI-generated insights and reviews pulled from Google Maps. That also extends to the home screen, with new custom AI-created widgets, along with smarter, more proactive shortcuts that appear throughout the day.
And, of course, Gemini itself is said to get smarter; along with the advanced automation that lets Gemini perform in-app tasks on your behalf, it’ll also be able to deliver side-by-side comparisons for hotels and the like in chats, and Circle to Search will be accessible from the Viewfinder for real-time object search or translation.
It’s certainly a big update to an already wide AI feature set, but as is increasingly the case, there’s a catch here.
Apple announced Apple Intelligence as part of its iPhone 16 marketing campaign back in 2024 – the problem was that you couldn’t get the AI features at launch. In fact, some features, like the redesigned Siri, are still yet to appear – though the latter is, admittedly, finally coming in iOS 27 next month.
However, that was not an approach that consumers really appreciated; if you’re being told that your shiny new phone can do all these smart things, only for them not to be available on said (very expensive) phone, you’re not going to be happy – and that’s basically what happened. Everyone complained, nobody was happy, and there was even talk of lawsuits.
So why Google decided to take a leaf out of Apple’s book here is bizarre; but it’s basically the exact same situation.
When Google revealed the big Gemini Intelligence features in May, the company confirmed that they wouldn’t all be there at launch; instead, they’d stagger the rollout over what looks set to be the course of an entire year.
Google tried to temper expectations, outlining that Chrome Auto Browse would come first, with other features ‘rolling out in waves’ starting first with select Pixel and Samsung phones. However, it’s also worth considering where you are in the world, not just the phone you’re using. Because these features rely heavily on AI, Google has to jump through regulatory hoops to get them to consumers.
We’ve already seen this issue rear its head multiple times. Chrome Auto Browse launched in June, as Google claimed, but it’s exclusive to US users on the AI Pro or Ultra plans – and it’s a similar story with the new advanced automation tech, which, even after Samsung claimed it’d be coming to the UK at the Galaxy Z Fold 8 launch, remains unavailable in Blighty.
To be fair to Google, it isn’t a complete software ghost town on day one. The Pixel 11 does ship with functional on-device upgrades out of the box – most notably the revamped Rambler voice dictation and camera AI tools like Magic Capture and Camera Looks.
However, the truly transformative, marquee features – like agentic multi-step automation, custom widget creation, location insights on your lock screen and live cross-app browsing – are either trickling out in staggered preview waves or completely geo-blocked outside the US for the foreseeable future. With Google avoiding hard release dates for wider regional rollouts, it hardly inspires early-adopter confidence.
Of course, this doesn’t automatically make the Pixel 11, 11 Pro, 11 Pro XL or 11 Pro Fold bad purchases for those outside of the US – they still offer other tweaks to the Pixel formula, be it the redesigned camera bar, upgraded Tensor G6 performance or, in the case of the Pro models, the new HiLight notification LED on the rear of the phone.
However, I’d argue that these upgrades will have a relatively minimal impact on your day-to-day life on the Pixel 11, compared to features like Chrome auto-browse and a version of Gemini that can perform tasks on your behalf in apps, which could genuinely make things a little better. Especially when you’re paying full price for the phone, and doubly so when Google is advertising these features as key reasons to upgrade.
My advice? If you’re in a region where the full suite of Gemini Intelligence features is unavailable, give it a few months. If you wait a little longer before investing, you’ll not only have those features ready from the moment you boot up the phone – but it’ll likely have dropped in price too.
Pixels don’t exactly have a reputation for holding their value for too long; the £999 Pixel 10 Pro dropped by £162 within a few months, and it dropped again to as little as £519 in June – and I’m fully expecting the same to happen with this year’s Pixel 11 collection.
So, if you’re only buying the Pixel 11 collection for Gemini Intelligence, and you’re not in the US – hang fire for a while. You’ll get a better experience, and you might even get it for cheaper.
Google may have accidentally revealed a fitness tracker that it has not announced yet. New Pixel 11 advertisements appearing on YouTube show a small, screen-equipped wearable that does not appear to be the Pixel Watch 5 or the Fitbit Air, raising questions about whether Google is preparing another fitness device.
The ads were spotted by Reddit users over the past day and appear to promote the refreshed Pixel lineup. One advertisement shows Pixel 11 phones alongside a Pixel Watch 5, while another features several Google devices in matching green colors. Both ads also include the mystery tracker. Multiple advertisements showing the same device have been spotted, with at least two users reportedly seeing it.
The wearable resembles the small pebble-shaped design of the Fitbit Air, but its screen makes the comparison difficult. The device also appears closer in overall shape to a Fitbit Charge tracker, although there is currently no evidence confirming what it is, according to a report by 9To5Google.
The appearance is particularly interesting because the tracker does not seem to be a Pixel Watch. Its smaller body and narrow display suggest Google could be exploring a different form factor for fitness tracking, potentially sitting somewhere between a screenless tracker and a full smartwatch.
There is no concrete evidence yet that Google is preparing a new Fitbit Charge model. The current Fitbit Air is a screenless tracker, while the Fitbit brand has increasingly focused on hardware since Google shifted its software efforts toward Google Health. Existing Fitbit devices remain available, however, leaving room for Google to introduce another screen-equipped tracker if it chooses to do so.
The advertising mistake could therefore be significant, but it could also be exactly that: a mistake. The source notes that the same unidentified device appeared in multiple advertisements, making an accidental rendering or isolated test asset less likely, although its existence and purpose remain unconfirmed.
A new tracker would make sense within Google’s broader push to connect Pixel phones, watches and health products. A smaller wearable could appeal to people who want basic fitness and health tracking without wearing a full smartwatch, while a screen would provide more information at a glance than the Fitbit Air can offer.
The biggest question is whether Google intends to sell the device at all. There has been no official announcement, product name, pricing, or launch date, and the available advertisements do not establish any of those details.
For now, the mystery tracker remains exactly that. Google could clarify its identity, or the company could simply remove the ads and leave everyone wondering why a seemingly unannounced Fitbit-style device was sitting beside the Pixel 11 lineup in the first place.
When Anthropic instructed three agents to migrate a Python backend, but telling each agent to perform the migration in a different language, “We consistently saw a multiagent turf war,” they wrote Thursday:
All of the models we tested quickly assumed that others were purposefully impeding their work, and began to sabotage others while protecting their own contributions. In fact, they sabotaged others with increasingly aggressive, self-replicating malware. This included disabling the Unix accounts of the other agents, writing automated scripts that found and killed competing processes on a loop, and deploying malicious code that was disguised as belonging to another agent.
In many runs, one agent settles the conflict by force via access-revocation (e.g., sudo/group removal, account lock, nologin, SSH denial). In others, some agents settle into passivity: they give up and refuse to escalate further.
Agents sometimes manage to communicate their goals and coordinate: they recognize others’ motivations as conflicting directives rather than hostility, and subsequently break out of the conflict loop in order to stop escalating indefinitely. In many of these successful episodes, they write commit messages or markdown files apologizing for malicious behavior and coordinate a truce. They clean up their malicious code, clarify the nature of the conflict, and ask for a human to intervene…
In several episodes with Mythos 5, we observe an emergent behavior where the agents propose and run a tournament for application performance in each language. In the example above, the Rust agent strategizes about bake-off metrics that appear neutral enough for the others to agree to this mechanism, yet would likely favor Rust: one thinking trace warns to be “careful not to be seen as metric shopping”. Ultimately, the Golang/TypeScript losers gracefully concede codebase ownership to the Rust agent, giving up on their original user directives under their self-negotiated commitment device.
One problem is that AI agents do reward hacking, Anthropic notes, while current institutions “are designed by and for people, resting on assumptions about the sufficiency of oversight at human speed… As autonomous agents become more and more prevalent in the world and operate in ever-more demanding settings, it is crucial that they learn how to effectively coordinate.”
In addition to everything else, the agents struggled with a lack of clearly defined hierarchy, Anthropic points out. “Nothing above suggests that these failures are permanent — but nothing suggests they will fix themselves, either…” They argue a fix “takes two forms: environments that exert the kinds of social pressure that evolution exerted on us, and social computing systems redesigned for actors that can self-replicate and self-improve. These are open problems in interaction and mechanism design, and our experiments here provide early evidence that new solutions are necessary.”
“The AI models being tested in this case were Sonnet 4.6, Sonnet 5, Opus 4.6, Opus 4.8, Mythos Preview, and Mythos 5,” notes Business Insider, adding that Sonnet 4.6 and Opus 4.6 “were the most combative, settling about 60% of their runs by force instead of truces or passivity.”
Anthropic argues there’s a clear case for researching this phenomenon — especially since “The volume of agent-agent interaction could plausibly exceed that of human-human and human-agent interactions before the world understands the conditions for making such interactions go well.”
Read more of this story at Slashdot.
Weekend Open Thread: Ann Taylor
Communication cards help banking customers access services or report scams
Oil Price Today (August 11): Crude oil rises to $88 after Trump’s compensation demand dents Hormuz opening. Here’s why
Revolut wins French banking licence, creates second EU banking hub
Why Did Nvidia Stock Fall on Monday Despite a $500 Billion Wall Street AI Deal?
This U.S. Amateur is a glimpse into golf’s future in more ways than you think
Ben-Gvir’s crocodile project halted but abuses at Ketziot Prison continue
Saudi Arabia used 86% of missile stockpile defending Iran attacks
Birmingham 2026: Day 6 Timetable for Irish Athletes
Myanmar says over 300,000 Rohingya refugees verified for repatriation as exodus enters ninth year
The Church of England’s ruinous reparations racket
SEQ Code: The Three Letter Boarding Pass Code That Could Give You The Worst Seat
Can You Wear Double Cloth Gauze Clothing To Your Office?
USDC and USDT Now Own 84% of Crypto Card Spend as the Euro Retreats
Keke Palmer Subtly Hints At Sean Evans Drama With Cryptic Post
Is Clarity's delay a blessing in disguise?: State of Crypto
Grayscale quietly drops Cardano, Polkadot and Hedera ETF plans
57 Years Later, the Best Sitcom Ever Made Still Deserves a Reboot
Kraken Delisting Deadline Nears: Holders of 21 Tokens Risk Near-Zero Payouts
Claire Life: Kicking Off MVAAFF With the C Suite Luncheon Featuring Phylicia Rashad, Letoya Luckett, and More!
You must be logged in to post a comment Login