Multimodal modeling focuses on building systems to understand and generate content across visual and textual formats. These models are designed to interpret visual scenes and produce new images using natural language prompts. With growing interest in bridging vision and language, researchers are working toward integrating image recognition and image generation capabilities into a unified system. This approach eliminates the need for separate pipelines and opens the path to more coherent and intelligent interactions across modalities. A key challenge in this field is to develop architectures that handle both understanding and generation without compromising the quality of either. Models need to grasp complex visual concepts and produce high-quality images matching user prompts. The difficulty lies in identifying suitable picture representations and training procedures that support both tasks. This problem becomes more evident when the same model is expected to interpret detailed text descriptions and generate visually accurate outputs based on them. It requires alignment of semantic understanding and pixel-level synthesis. Previous approaches have generally used Variational Autoencoders (VAEs) or CLIP-based encoders to represent images. VAEs are efficient for reconstruction but encode lower-level features, often leading to less informative representations. CLIP-based encoders provide high-level semantic embeddings by learning from large-scale image-text pairs. However, CLIP was not built for image reconstruction, making it challenging to use for generation unless paired with models like diffusion decoders. In terms of training, Mean Squared Error (MSE) is widely used for simplicity but tends to produce deterministic outputs. To improve generation diversity and quality, researchers have turned to Flow Matching, which introduces controlled stochasticity and better models the continuous nature of image features. Researchers from Salesforce Research, in collaboration with the University of Maryland and several academic institutions, introduced BLIP3-o, a family of unified multimodal models. The model adopts a dual-stage training strategy where image understanding is learned first, followed by image generation. The proposed system leverages CLIP embeddings to represent images and integrates them with a diffusion transformer to synthesize new visual outputs. Unlike previous joint training methods, the sequential approach maintains the strength of each task independently. The diffusion module is trained while keeping the autoregressive backbone frozen, avoiding task interference. To improve alignment and visual fidelity, the team also curated BLIP3o-60k, a high-quality instruction-tuning dataset created by prompting GPT-4o across varied visual categories, including scenes, objects, gestures, and text. They developed two model versions: an 8-billion parameter model trained with proprietary and public data, and a 4-billion version using only open-source data. The image generation pipeline of BLIP3-o is built on Qwen2.5-VL large language models. Prompts are processed to produce visual features refined through a Flow Matching diffusion transformer. This transformer is based on the Lumina-Next architecture, optimized for speed and quality with 3D rotary position embedding and grouped-query attention. The model encodes each image into 64 fixed-length semantic vectors, regardless of resolution, which supports compact storage and efficient decoding. The research team used a large-scale dataset of 25 million images from sources like CC12M, SA-1B, and JourneyDB to train the models. They extended it with 30 million proprietary samples for the 8B model. They also included 60k instruction-tuning samples covering challenging prompts such as complex gestures and landmarks, generated via GPT-4o. In terms of performance, BLIP3-o demonstrated top scores across multiple benchmarks. The 8B model achieved a GenEval score of 0.84 for image generation alignment and a WISE score of 0.62 for reasoning ability. Image understanding scored 1682.6 on MME-Perception, 647.1 on MME-Cognition, 50.6 on MMMU, and 83.1 on both VQAv2 and TextVQA datasets. A human evaluation comparing BLIP3-o 8B with Janus Pro 7B showed that BLIP3-o was preferred 50.4% of the time for visual quality and 51.5% for prompt alignment. These results are supported by statistically significant p-values (5.05e-06 and 1.16e-05), indicating the superiority of BLIP3-o in subjective quality assessments. This research outlines a clear solution to the dual challenge of image understanding and generation. CLIP embeddings, Flow Matching, and a sequential training strategy demonstrate how the problem can be approached methodically. The BLIP3-o model delivers state-of-the-art results and introduces an efficient and open approach to unified multimodal modeling. Check out the Paper, GitHub Page and Model on Hugging Face. All credit for this research goes to the researchers of this project. Also, feel free to follow us on Twitter and don’t forget to join our 90k+ ML SubReddit. NikhilNikhil is an intern consultant at Marktechpost. He is pursuing an integrated dual degree in Materials at the Indian Institute of Technology, Kharagpur. Nikhil is an AI/ML enthusiast who is always researching applications in fields like biomaterials and biomedical science. With a strong background in Material Science, he is exploring new advancements and creating opportunities to contribute.Nikhilhttps://www.marktechpost.com/author/nikhil0980/Georgia Tech and Stanford Researchers Introduce MLE-Dojo: A Gym-Style Framework Designed for Training, Evaluating, and Benchmarking Autonomous Machine Learning Engineering (MLE) AgentsNikhilhttps://www.marktechpost.com/author/nikhil0980/This AI Paper Investigates Test-Time Scaling of English-Centric RLMs for Enhanced Multilingual Reasoning and Domain GeneralizationNikhilhttps://www.marktechpost.com/author/nikhil0980/PwC Releases Executive Guide on Agentic AI: A Strategic Blueprint for Deploying Autonomous Multi-Agent Systems in the EnterpriseNikhilhttps://www.marktechpost.com/author/nikhil0980/Multimodal AI Needs More Than Modality Support: Researchers Propose General-Level and General-Bench to Evaluate True Synergy in Generalist Models

Bungie has apologized for the uncredited, uncompensated use of another artist's work in the texture sheets for the Marathon alpha test.In a post on social media, the company publicly confirmed that a "former artist at Bungie" included art assets made by graphic designer "Antireal" (real name unknown). Antireal had accused the company of lifting her work in a post on Bluesky to decorate several in-game posters.This is a black mark on Bungie's record, especially for a company that's apologized for using assets lifted from other artists in the past. Using art assets the company doesn't have the rights to can deprive artists of badly-needed funds and put the studio at risk for legal exposure."Bungie is of course not obligated to hire me when making a game that draws overwhelmingly from the same design language I have refined for the last decade, but clearly my work was good enough to pillage for ideas and plaster all over their game without pay or attribution," claimed Antireal. She explained that the assets perfectly mirrored work they posted in 2017.Antireal's claims were bolstered by Marathon players who noticed a key fingerprint from Antireal's work was directly visible in the game. In 2024 the artist said they hit the "loss" meme in their work 6 years ago (the meme refers to recreations of a webcomic page from Ctrl+Alt+Delete with the same name). A Twitter user spotted the meme on an in-game building in Marathon.Related:Image via X."We immediately investigated a concern regarding unauthorized use of artist decals in Marathon and confirmed that a former Bungie artist included these in a texture sheet that was ultimately used in-game," Bungie said in a post confirming Antireal's claims. "This issue was unknown by our existing art team, and we are still reviewing how this oversight occurred. We take matters like this very seriously. We have reached out to [Antireal] to discuss this issue and are committed to do right by the artist."The company added that it is conducting a review of the former artist's work and implementing "stricter checks" on its art team to ensure no inappropriate assets are used.Stolen art keeps appearing in Bungie's games"We value the creativity and dedication of all artists who contribute to our games, and we are committed to doing right by them," Bungie stated in its post confirming the use of Antireal's art. There's no reason to doubt that claim—but this isn't the first time Bungie has admitted to stolen versions of other artist's work appearing in its games.Related:In 2024, it confirmed a piece of "fan art" was used to decorate an in-game collectible in Destiny 2. In 2023 it said it would "compensate and credit" an artist whose work was used in a cutscene produced in part by an "external vendor." And in 2021 principal community manager Chris Shannon apologized to a fan artist in another trailer.Bungie's also found itself in hot water over other art assets. In 2017 it apologized and removed an in-game item that represented a real-world white supremacist flag.Apologizing and acknowledging the wrongdoing is one way to address what seems to be unintentional plagiarism, however that approach doesn't undo the harm done to Antireal and other artists who find themselves victims of plagiarism, intentional or no."In 10 years I have never made a consistent income from this work," Antireal said. "I am tired of designers from huge companies moodboarding and parasitising my designs while I struggle to make a living."

The Ontario Science Centre was shuttered on June 21, 2024, with the Ontario government claiming that a structural roof issue made it unsafe. New documents suggest that the engineers inspecting the roof didn’t agree. Photo by Canmenwalker via Wikimedia Commons New documents, obtained by Canadian Architect through a freedom of information request, appear to indicate that Rimkus—the structural engineers hired to assess the Ontario Science Centre’s roof before its closure—did not support closing the Centre for public safety reasons. The draft versions of the report instead recommend regular, routine maintenance to the roofs over the course of the coming 20 years. When read in conjunction with correspondence preceding the closure obtained by Global News, it seems likely that the engineers were pressured by the government to produce a report that would support the shuttering of the Ontario Science Centre. However, such a recommendation was not supported by their research and expertise. Canadian Architect obtained draft versions of parts of the Rimkus report from March, April, and May of 2024. The report, whose final version is dated June 18, 2024, was commissioned to assess the condition of the Science Centre’s reinforced autoclaved aerated concrete (RAAC) roof panels, an outdated form of construction that can pose problems over time if not properly maintained. It was based on on-site investigations conducted by Rimkus from December 2023 to March 2024. In a key section on RAAC Panel Recommendations, the March and April versions of the report clearly recommend to proceed with repairs to a set of panels identified as “high risk”—less than 2.5% of the Science Centre’s roofs—over the summer. It reads: “It is Rimkus’ recommendation to address all 2024 RAAC panel repair (amber risk) locations as a single project. As such, it is our opinion that the most cost effective strategy would be a consistent repair approach. Panel replacement rather than reinforcement is feasible at all areas, and presented in Table 7.” The government seems to have pushed the engineers to go further, asking how the RAAC risk could be entirely eliminated. The May version adds the sentences: “It is recommended that RAAC panels should be replaced at the time of the next scheduled roof replacement, to completely eliminate the RAAC panel risk. Recommended timelines for complete roof assembly and RAAC panel (where applicable) replacements, and budgetary costs, by roof area are shown in Table 8.” In the March, April, and May versions of the report, there is no provision for declining to do the repairs—simply because roof repairs are a relatively routine, straightforward undertaking. RAAC panels at the Science Centre had been routinely replaced in the past. At the time of the inspection, a half dozen panels were identified as being in critical condition, and were already being replaced while the report was completed. A building permit was in progress to repair a section of the roof over the rainforest room. The cost of the immediate repairs needed would amount to around $500,000—or just over $7 million, if coupled with replacing, repairing, and maintaining larger sections of the Science Centre’s roofs that year. And yet, the language apparently wasn’t strong enough for the Ontario government, whose communications staffers aimed to justify the closure as a public safety issue. For the final version of the report, government officials seem to have pressed the engineers to answer the question: what if we just don’t do the repairs? Wouldn’t the only totally safe option, in that case, be to close the building? As a result, the final version of the report included the following passage: “Where replacement or reinforcement cannot be completed within the recommended time frame, then one of the following supplemental risk mitigation options is recommended and should be implemented in conjunction with the snow/load monitoring program: Option 1: Restricted access or full closure to prevent any persons from walking in areas where high risk panels are present. Option 2: Installation of temporary shoring (reinforcement) supporting the underside of RAAC panels. Options 3: If shoring is not possible, installation of horizontal hoarding near the underside of hard ceiling levels, or other building interferences (sprinkler mains/process piping etc.) The above three recommended options are listed in order of preference, with option 1 completing [sic] eliminating the risk to public or staff.” As I have written before, while a quick reading of this passage may seem to suggest “full closure” of the Science Centre, a closer look shows that access would only need to be restricted at the areas “where high risk panels are present.” The high-risk panels represent a miniscule portion of the Science Centre’s roof, almost entirely in non-exhibition areas. Texts between Infrastructure Ontario minister Michael Lindsey and a top communications staffer, obtained by Global News, point to the Ministry’s communications with structural engineers Rimkus in the lead-up to the announcement of the Ontario Science Centre’s closure. The texts also show that the Ministry of Infrastructure was in tight communication with the Premier’s Office (PO) about the closure.   Text messages between Infrastructure Ontario CEO Michael Lindsay and the agency’s top communications staffer, from the run-up to the closure announcement, suggest that the agency was working closely with Rimkus and pressing executives to tailor its language before the report was submitted. “The executives at Rimkus look like survivors of the apocalypse. [laugh/cry emoji],” one of the texts reads. The presence of a typo in the report’s added section also suggests a rushed addition. When the final report was sent, a member of Rimkus’s team seems to have also referenced conversations that took place between the two parties. “Please refer to revised final summary report (R2) attached, with updated notes/paragraphs, as discussed,” they wrote on June 18. While some back-and-forth between consultants and public clients is expected, it is unusual that this contact would be with the Ministry’s top brass—particularly for a routine, technical report. Also notable is that while the Rimkus report released to the public by Infrastructure Ontario was signed by the engineers, it was not stamped—an indication that the engineers may not have felt entirely comfortable with how the report landed, and were unwilling to be professionally liable for it. In an email to top Ministry of Infrastructure and Infrastructure Ontario officials after the closure announcement, an Ontario Ministry director of communications shares a few questions that may arise. One is: “Q: Is this [closure] just a convenient excuse to move [the Ontario Science Centre] to Ont Place?” He suggests the reply: “This is a health and safety issue. Full stop.” But for the engineers that actually inspected the roof, closure should never have been on the table: the roof issue could—and should—have been addressed with routine repairs. The post New documents reinforce that Science Centre closure was not supported by engineers appeared first on Canadian Architect.

The Internet has greatly shortened product life cycles. Objects that stood the test of time offer lessons for today. So do designs that are intentionally ephemeral.Designing applications or websites means your designs don’t survive very long. Even if your product or app survives for a decade or more, there will be pressure to update the design every few years. This got me wondering how long designs of other things stay around. So here is a curated survey of how long our designs survive.Sticky NotesThe sticky note… a lab experiment gone awry(average lifespan: several hours) — The ubiquitous sticky note famously emerged from a lab experiment gone awry. Glue that wasn’t very sticky. Although sticky notes have been around for 45 years, the life of any single sticky note is probably measured in hours, not days. Nevertheless they have been a boon to brainstorming sessions and design sprints around the world. Their ephemeral nature is the key to their value. Don’t like what you wrote? Throw it out and write another one. Don’t think it belongs in this group? Move it to another. Sticky notes are the rare object that is as easily used digitally as their physical counterparts.Sand MandalaA finished sand mandala — Attribution: By the original uploader was Colonel Warden at English Wikipedia. — Transferred from en.wikipedia to Commons by SMasters using CommonsHelper., CC BY-SA 3.0, https://commons.wikimedia.org/w/index.php?curid=16371236(in use for a few days or weeks) — A sand mandala is a symbolic representation of the universe made of colored sand. Creating them is a ceremonial act practiced by Tibetan monks and nuns and takes a few days to several weeks to create. Constructing one is an act of meditation requiring immense focus. After the mandala is finished, it is on view for only a short period of time. It is then destroyed in a dissolution ceremony meant to emphasize the impermanence of all things. Finally the sand is released into a river to transfer the spirits embodied in the mandala back into nature and the rest of the world.The first time I saw a video of this process, it shocked my world. The design was vivid, ornate, and beautiful. To see something so beautiful then destroyed by the people who made it jolted me. When swept into a pile, the vivid pigments quickly become a single gray mass. They are an amazing lesson about the nature of our lives and the world of things.WebsitesThe SF bay area edition of Craigslist.com barely changed from when it launched in 1995!(in use for 32 years) — The lifespan of your average website is about 2 years according to this post. Many sites are up for much longer but even those tend to need a refresh. Craigslist is the notable exception by still proudly sporting its original 1995 web 1.0 look and feel. I’ve worked on around 14 websites and apps. Of those, only 2 (7%) are live and still reflect the design I worked on. As dominant as the Internet is in today’s culture, “built to last” simply does not apply. Permanence is an anti-pattern.Thonet №14 ChairThe Thonet №14 Chair — Attribution: This file is licensed under the Creative Commons Attribution-Share Alike 3.0 Unported license.(in use for 166 years) — Michael Thonet’s №14 Chair is a design classic. Simple and minimal. Introduced in 1859, It was the first chair to be mass produced but does not look like an industrial product. It is sometimes called the cafe chair due to its strong association with Parisian cafes. It was also sometimes referred to as the “lion-tamer chair”. You can see one in use in this post.BicycleThe ‘Flying Pigeon’ bicycle. Since the 1950s, over 500 million of these bikes have been produced and sold.(in use for 208 years) — The bicycle was invented in Germany and introduced in Paris around 1817. It is an amazing design object matched only by its engineering simplicity. The basic triangular geometry is minimal, strong, and still recognizable 200 years after being invented. There have been many improvements to the underlying technologies but the basic design remains.BuildingsThe Pantheon of Rome — Attribution: By Macrons — Own work, CC BY-SA 4.0, https://commons.wikimedia.org/w/index.php?curid=49683083The Pantheon in Rome was completed around 126 AD and has been in continuous use for almost 2000 years. It is widely considered the oldest building still in continuous use today. It began as a Roman temple. For the ceiling of the Parthenon, the Romans used cement mixed with ash which kept it light but strong. In the 7th century, it became a Christian church, the Basilica of St. Mary and the Martyrs. Adaptive reuse helped preserve it through centuries of war, weather, and regime changes.“Built to last” is a phrase sometimes used for things like old cars, buildings, and walls. It certainly applies to the pyramids and the Pantheon. It is much less frequently applied to objects today. It doen’t apply to the majority of objects we use in our daily lives. Consumer culture has driven us towards a throw away culture. This impacts the environment and our ability to build things to last.What does it take to build something that will last?This is a question asked by Louis Kahn, one of the great American architects. His Exeter Library is his answer. He wanted to pursue an architecture that would stand the test of time. He wanted to build something that would be monumental, strong and, like the pyramids, remain beautiful even in ruin.The interior of the Phillips Exeter Academy Library showing the raw, monumental cement building structure. Image source: https://www.reddit.com/r/ArchitecturePornCities(in use for 11,000 years) — The title of oldest continuously inhabited city is a subject of debate but Damascus is often credited as the oldest city still in use today. Jericho is older but was not in continuous use. Jerusalem & Athens both date back about 5000 years. From a design perspective, I marvel at Venice. It’s been in existence for over 1000 years. Its lasting beauty is amazing when you realize it was originally built to escape invaders on the mainland. The buildings, plazas, along with the canals and bridges that connect them are an amazing experience.The Bucentaur Returns to the Pier at the Doge’s Palace (c. 1730) by Canaletto — Attribution: Google Arts & Culture — mwEV7sO9uSFCpw, Public Domain, https://commons.wikimedia.org/w/index.php?curid=22572210Mortar & PestleThe mortar & pestle — symbol of the pharmacist — Attribution: Image by Evan-Amos — Own work, Public Domain, https://commons.wikimedia.org/w/index.php?curid=12038229Beyond using our teeth, we’ve been grinding up food and other materials for about 37,000 years. The Mortar & Pestle was likely used initially to prepare grains and seeds to be more digestible but they had many other uses.They are still widely used today (guacamole anyone?) and have come to symbolize the pharmaceutical profession. As someone who cooks, I’m glad we figured out ground spices makes tastier dishes.Pyramids(built 4000 years ago) — The Great Pyramid of Giza was the tallest human-made structure in the world for 3800 years. I’m including pyramids for age reference only. How long they’ve been “in use” raises tricky questions. Should we measure use by the initial construction; as amazing ruins we only look at; or as a tomb for a dead pharaoh? They certainly were amazing feats of engineering.Bow & arrow / stone arrowheads(in use for 37,000–74,000 years) — These two designs are best taken as a connected journey. Arrows are an evolution of the stone-tipped spear. The bow & arrow was a leap in the efficiency and precision of throwing a spear. The innovation allowed hunters to remain farther away from their prey. The bow, arrow, and stone arrowhead built on the previous technology (an interesting concept known as “the adjacent possible” richly described by Steven Johnson in, ‘Where Good Ideas Come From’). Each element required an increase in precision and skill and offered advances over the previous tool. Older arrowheads were mostly focused on the general shape and cutting edges. Later versions were shaped to be more firmly lashed to wooden spears. As the first to enter the battlefield, U.S. Marines sometimes refer to themselves as, “the tip of the spear”. A 74,000 year old metaphor that is still understandable today.Hand axeIllustration of a hand axe.Imagine designing a product that was so good it was used for a million years. That’s 1,000,000 years! I remember seeing a hand axe in the “Tools: Extending Our Reach” exhibit at the Smithsonian’s Cooper Hewitt National Design Museum. I pondered it as a design object that dwarfed my career as a designer. The hand axe is a physical testament to both the slowness of early hominid evolution and the enduring utility of the object. The key innovation was the ever more skilled chipping away of flakes to create the cutting edge. Older versions of the axe are barely recognizable as being intentionally created. Later versions however, show two distinct functional requirements. The part that you held needed to be broad and round so as to be comfortable in the hand. While “the business end” had sharp, carefully-shaped cutting edges. I marvel at anything being used for this incomprehensible span of time.Final thoughts“If you want to go quickly, go alone. But if you want to go far, go together.” — African ProverbAs I wrote this post, I thought about the lessons of these very different objects. Here are my takeaways.Monumental acts take monumental teams — It takes incredible effort by many people to build something that has great impact and will last. Monumental achievements are never accomplished by one person. Each hand axe was likely created by one person but the concept we now know as the hand axe happened over a million years of usage. Venice is a singular experience. But building it required the work of many people over centuries.Adaptation is key to survival — The Pantheon survived in part because it is a space flexible in use. If your designs are to survive, you must hand them off to others. Loving caretakers are as important as passionate creators.Celebrate the ephemeral — Sand mandalas and sticky notes allow us to create something without becoming too attached to it. Having worked on lots of long-term projects, I’ve always found making dinner to be a positive antidote. Chopping vegetables can be meditative. A meal is finished in an hour or 2 and then you get to enjoy the creation. Simple acts are still creative.A graph showing how long these designs have been use. The hand axe was excluded because its million years of use reduces the other objects to tiny, barely visible bars.ReferenceTimeline of oldest human inventions — WikipediaOldest still standing buildings — WikipediaHow long designs survive was originally published in UX Collective on Medium, where people are continuing the conversation by highlighting and responding to this story.

Bong Joon-ho’s follow-up to Parasite is coming to streaming on Max in just a few days.Mickey 17, starring Robert Pattinson as multiple versions of a menial worker in a dark future, is the latest film from the Academy Award-winning Bong. Pattinson’s Mickey agrees to become an “expendable” in order to escape a debt on Earth. He’s assigned to a colonial mission to outer space, where he is used to accomplish all the most dangerous tasks; every time he dies, he’s given a new cloned body and put back to work.He’s Mickey 17 because he’s now on his 17th body — at least until he seemingly dies in another work-related accident and the colonists create a Mickey 18. But the first Mickey somehow survives, leading to two Mickeys (which is against the law in this bleak future) running around their space ship.In addition to Pattinson in the two lead roles, the film also stars Naomi Ackie, Steven Yeun, Toni Collette, and Mark Ruffalo as Kenneth Marshall, the former politician turned quasi cult leader who is heading the expedition into space to find a new planet capable of sustaining human life.Warner Bros.Warner Bros.loading...READ MORE: 10 Horror Movies So Extreme They Made Audiences SickHere is the film’s official synopsis:From the Academy Award®-winning writer/director of “Parasite,” Bong Joon Ho, comes his next groundbreaking cinematic experience, MICKEY 17. The unlikely hero, Mickey Barnes (Robert Pattinson) has found himself in the extraordinary circumstance of working for an employer who demands the ultimate commitment to the job… to die, for a living.Mickey 17 premieres on Max (soon to be HBO Max again) on May 23. It premieres on HBO proper on Saturday, May 24 at 7:40 PM.Get our free mobile appThe Best Box-Office Bombs in HistoryThese films flopped in theaters — but they are worth watching now.Filed Under: Bong Joon-ho, Max, Mickey 17, Robert PattinsonCategories: Movie News

May 16, 2025Ravie LakshmananMalware / Cyber Attack Cybersecurity researchers have shed light on a new malware campaign that makes use of a PowerShell-based shellcode loader to deploy a remote access trojan called Remcos RAT. "Threat actors delivered malicious LNK files embedded within ZIP archives, often disguised as Office documents," Qualys security researcher Akshay Thorve said in a technical report. "The attack chain leverages mshta.exe for proxy execution during the initial stage." The latest wave of attacks, as detailed by Qualys, employs tax-related lures to entice users into opening a malicious ZIP archive containing a Windows shortcut (LNK) file, which, in turn, makes use of mshta.exe, a legitimate Microsoft tool used to run HTML Applications (HTA). The binary is used to execute an obfuscated HTA file named "xlab22.hta" hosted on a remote server, which incorporates Visual Basic Script code to download a PowerShell script, a decoy PDF, and another HTA file similar to xlab22.hta called "311.hta." The HTA file is also configured to make Windows Registry modifications to ensure that "311.hta" is automatically launched upon system startup. Once the PowerShell script is executed, it decodes and reconstructs a shellcode loader that ultimately proceeds to launch the Remcos RAT payload entirely in memory. Remcos RAT is a well-known malware that offers threat actors full control over compromised systems, making it an ideal tool for cyber espionage and data theft. A 32-bit binary compiled using Visual Studio C++ 8, it features a modular structure and can gather system metadata, log keystrokes, capture screenshots, monitor clipboard data, and retrieve a list of all installed programs and running processes. In addition, it establishes a TLS connection to a command-and-control (C2) server at "readysteaurants[.]com," maintaining a persistent channel for data exfiltration and control. This is not the first time fileless versions of Remcos RAT have been spotted in the wild. In November 2024, Fortinet FortiGuard Labs detailed a phishing campaign that filelessly deployed the malware by making use of order-themed lures. What makes the attack method attractive to threat actors is that it allows them to operate undetected by many traditional security solutions as the malicious code runs directly in the computer's memory, leaving very few traces on the disk. "The rise of PowerShell-based attacks like the new Remcos RAT variant demonstrates how threat actors are evolving to evade traditional security measures," J Stephen Kowski, Field CTO at SlashNext, said. "This fileless malware operates directly in memory, using LNK files and MSHTA.exe to execute obfuscated PowerShell scripts that can bypass conventional defenses. Advanced email security that can detect and block malicious LNK attachments before they reach users is crucial, as is real-time scanning of PowerShell commands for suspicious behaviors." The disclosure comes as Palo Alto Networks Unit 42 and Threatray detailed a new .NET loader that's used to detonate a wide range of commodity information stealers and RATS like Agent Tesla, NovaStealer, Remcos RAT, VIPKeylogger, XLoader, and XWorm. The loader features three stages that work in tandem to deploy the final-stage payload: A .NET executable that embeds the second and third stages in encrypted form, a .NET DLL that decrypts and loads the next stage, and a .NET DLL that manages the deployment of the main malware. "While earlier versions embedded the second stage as a hardcoded string, more recent versions use a bitmap resource," Threatray said. "The first stage extracts and decrypts this data, then executes it in memory to launch the second stage." Unit 42 described the use of bitmap resources to conceal malicious payloads a a steganography technique that can bypass traditional security mechanisms and evade detection. The findings also coincide with the emergence of several phishing and social engineering campaigns that are engineered for credential theft and malware delivery - Use of trojanized versions of the KeePass password management software – codenamed KeeLoader – to drop a Cobalt Strike beacon and steal sensitive KeePass database data, including administrative credentials. The malicious installers are hosted on KeePass typosquat domains that are served via Bing ads. Use of ClickFix lures and URLs embedded within PDF documents and a series of intermediary dropper URLs to deploy Lumma Stealer. Use of booby-trapped Microsoft Office documents that are used to deploy the Formbook information stealer protected using a malware distribution service referred to as Horus Protector. Use of blob URIs to locally loads a credential phishing page via phishing emails, with the blob URIs served using allow-listed pages (e.g., onedrive.live[.]com) that are abused to redirect victims to a malicious site that contains a link to a threat actor-controlled HTML page. Use of RAR archives masquerading as setup files to distribute NetSupport RAT in attacks targeting Ukraine and Poland. Use of phishing emails to distribute HTML attachments that contain malicious code to capture victims' Outlook, Hotmail, and Gmail credentials and exfiltrate them to a Telegram bot named "Blessed logs" that has been active since February 2025 The developments have also been complemented by the rise in artificial intelligence (AI)-powered campaigns that leverage polymorphic tricks that mutate in real-time to sidestep detection efforts. These include modifying email subject lines, sender names, and body content to slip past signature-based detection. "AI gave threat actors the power to automate malware development, scale attacks across industries, and personalize phishing messages with surgical precision," Cofense said. "These evolving threats are increasingly able to bypass traditional email filters, highlighting the failure of perimeter-only defenses and the need for post-delivery detection. It also enabled them to outmaneuver traditional defenses through polymorphic phishing campaigns that shift content on the fly. The result: deceptive messages that are increasingly difficult to detect and even harder to stop." Found this article interesting? Follow us on Twitter  and LinkedIn to read more exclusive content we post. SHARE    

I’ve been working as a data science consultant for the past three years, and I’ve had the opportunity to work on multiple projects across various industries. Yet, I noticed one common denominator among most of the clients I worked with: They rarely have a clear idea of the project objective. This is one of the main obstacles data scientists face, especially now that Gen AI is taking over every domain. But let’s suppose that after some back and forth, the objective becomes clear. We managed to pin down a specific question to answer. For example: I want to classify my customers into two groups according to their probability to churn: “high likelihood to churn” and “low likelihood to churn” Well, now what? Easy, let’s start building some models! Wrong! If having a clear objective is rare, having a reliable benchmark is even rarer. In my opinion, one of the most important steps in delivering a data science project is defining and agreeing on a set of benchmarks with the client. In this blog post, I’ll explain: What a benchmark is, Why it is important to have a benchmark, How I would build one using an example scenario and Some potential drawbacks to keep in mind What is a benchmark? A benchmark is a standardized way to evaluate the performance of a model. It provides a reference point against which new models can be compared. A benchmark needs two key components to be considered complete: A set of metrics to evaluate the performance A set of simple models to use as baselines The concept at its core is simple: every time I develop a new model I compare it against both previous versions and the baseline models. This ensures improvements are real and tracked. It is essential to understand that this baseline shouldn’t be model or dataset-specific, but rather business-case-specific. It should be a general benchmark for a given business case. If I encounter a new dataset, with the same business objective, this benchmark should be a reliable reference point. Why building a benchmark is important Now that we’ve defined what a benchmark is, let’s dive into why I believe it’s worth spending an extra project week on the development of a strong benchmark. Without a Benchmark you’re aiming for perfection — If you are working without a clear reference point any result will lose meaning. “My model has a MAE of 30.000” Is that good? IDK! Maybe with a simple mean you would get a MAE of 25.000. By comparing your model to a baseline, you can measure both performance and improvement. Improves Communicating with Clients — Clients and business teams might not immediately understand the standard output of a model. However, by engaging them with simple baselines from the start, it becomes easier to demonstrate improvements later. In many cases benchmarks could come directly from the business in different shapes or forms. Helps in Model Selection — A benchmark gives a starting point to compare multiple models fairly. Without it, you might waste time testing models that aren’t worth considering. Model Drift Detection and Monitoring — Models can degrade over time. By having a benchmark you might be able to intercept drifts early by comparing new model outputs against past benchmarks and baselines. Consistency Between Different Datasets — Datasets evolve. By having a fixed set of metrics and models you ensure that performance comparisons remain valid over time. With a clear benchmark, every step in the model development will provide immediate feedback, making the whole process more intentional and data-driven. How I would build a benchmark I hope I’ve convinced you of the importance of having a benchmark. Now, let’s actually build one. Let’s start from the business question we presented at the very beginning of this blog post: I want to classify my customers into two groups according to their probability to churn: “high likelihood to churn” and “low likelihood to churn” For simplicity, I’ll assume no additional business constraints, but in real-world scenarios, constraints often exist. For this example, I am using this dataset (CC0: Public Domain). The data contains some attributes from a company’s customer base (e.g., age, sex, number of products, …) along with their churn status. Now that we have something to work on let’s build the benchmark: 1. Defining the metrics We are dealing with a churn use case, in particular, this is a binary classification problem. Thus the main metrics that we could use are: Precision — Percentage of correctly predicted churners among all predicted churners Recall — Percentage of actual churners correctly identified F1 score — Balances precision and recall True Positives, False Positives, True Negative and False Negatives These are some of the “simple” metrics that could be used to evaluate the output of a model. However, it is not an exhaustive list, standard metrics aren’t always enough. In many use cases, it might be useful to build custom metrics. Let’s assume that in our business case the customers labeled as “high likelihood to churn” are offered a discount. This creates: A cost ($250) when offering the discount to a non-churning customer A profit ($1000) when retaining a churning customer Following on this definition we can build a custom metric that will be crucial in our scenario: # Defining the business case-specific reference metric def financial_gain(y_true, y_pred): loss_from_fp = np.sum(np.logical_and(y_pred == 1, y_true == 0)) * 250 gain_from_tp = np.sum(np.logical_and(y_pred == 1, y_true == 1)) * 1000 return gain_from_tp - loss_from_fp When you are building business-driven metrics these are usually the most relevant. Such metrics could take any shape or form: Financial goals, minimum requirements, percentage of coverage and more. 2. Defining the benchmarks Now that we’ve defined our metrics, we can define a set of baseline models to be used as a reference. In this phase, you should define a list of simple-to-implement model in their simplest possible setup. There is no reason at this state to spend time and resources on the optimization of these models, my mindset is: If I had 15 minutes, how would I implement this model? In later phases of the model, you can add mode baseline models as the project proceeds. In this case, I will use the following models: Random Model — Assigns labels randomly Majority Model — Always predicts the most frequent class Simple XGB Simple KNN import numpy as np import xgboost as xgb from sklearn.neighbors import KNeighborsClassifier class BinaryMean(): @staticmethod def run_benchmark(df_train, df_test): np.random.seed(21) return np.random.choice(a=[1, 0], size=len(df_test), p=[df_train['y'].mean(), 1 - df_train['y'].mean()]) class SimpleXbg(): @staticmethod def run_benchmark(df_train, df_test): model = xgb.XGBClassifier() model.fit(df_train.select_dtypes(include=np.number).drop(columns='y'), df_train['y']) return model.predict(df_test.select_dtypes(include=np.number).drop(columns='y')) class MajorityClass(): @staticmethod def run_benchmark(df_train, df_test): majority_class = df_train['y'].mode()[0] return np.full(len(df_test), majority_class) class SimpleKNN(): @staticmethod def run_benchmark(df_train, df_test): model = KNeighborsClassifier() model.fit(df_train.select_dtypes(include=np.number).drop(columns='y'), df_train['y']) return model.predict(df_test.select_dtypes(include=np.number).drop(columns='y')) Again, as in the case of the metrics, we can build custom benchmarks. Let’s assume that in our business case the the marketing team contacts every client who’s: Over 50 y/o and That is not active anymore Following this rule we can build this model: # Defining the business case-specific benchmark class BusinessBenchmark(): @staticmethod def run_benchmark(df_train, df_test): df = df_test.copy() df.loc[:,'y_hat'] = 0 df.loc[(df['IsActiveMember'] == 0) & (df['Age'] >= 50), 'y_hat'] = 1 return df['y_hat'] Running the benchmark To run the benchmark I will use the following class. The entry point is the method compare_with_benchmark() that, given a prediction, runs all the models and calculates all the metrics. import numpy as np class ChurnBinaryBenchmark(): def __init__( self, metrics = [], benchmark_models = [], ): self.metrics = metrics self.benchmark_models = benchmark_models def compare_pred_with_benchmark( self, df_train, df_test, my_predictions, ): output_metrics = { 'Prediction': self._calculate_metrics(df_test['y'], my_predictions) } dct_benchmarks = {} for model in self.benchmark_models: dct_benchmarks[model.__name__] = model.run_benchmark(df_train = df_train, df_test = df_test) output_metrics[f'Benchmark - {model.__name__}'] = self._calculate_metrics(df_test['y'], dct_benchmarks[model.__name__]) return output_metrics def _calculate_metrics(self, y_true, y_pred): return {getattr(func, '__name__', 'Unknown') : func(y_true = y_true, y_pred = y_pred) for func in self.metrics} Now all we need is a prediction. For this example, I made a quick feature engineering and some hyperparameter tuning. The last step is just to run the benchmark: binary_benchmark = ChurnBinaryBenchmark( metrics=[f1_score, precision_score, recall_score, tp, tn, fp, fn, financial_gain], benchmark_models=[BinaryMean, SimpleXbg, MajorityClass, SimpleKNN, BusinessBenchmark] ) res = binary_benchmark.compare_pred_with_benchmark( df_train=df_train, df_test=df_test, my_predictions=preds, ) pd.DataFrame(res) Benchmark metrics comparison | Image by Author This generates a comparison table of all models across all metrics. Using this table, it is possible to draw concrete conclusions on the model’s predictions and make informed decisions on the following steps of the process. Some drawbacks As we’ve seen there are plenty of reasons why it is useful to have a benchmark. However, even though benchmarks are incredibly useful, there are some pitfalls to watch out for: Non-Informative Benchmark — When the metrics or models are poorly defined the marginal impact of having a benchmark decreases. Always define meaningful baselines. Misinterpretation by Stakeholders — Communication with the client is essential, it is important to state clearly what the metrics are measuring. The best model might not be the best on all the defined metrics. Overfitting to the Benchmark — You might end up trying to create features that are too specific, that might beat the benchmark, but do not generalize well in prediction. Don’t focus on beating the benchmark, but on creating the best solution possible to the problem. Change of Objective — Objectives defined might change, due to miscommunication or changes in plans. Keep your benchmark flexible so it can adapt when needed. Final thoughts Benchmarks provide clarity, ensure improvements are measurable, and create a shared reference point between data scientists and clients. They help avoid the trap of assuming a model is performing well without proof and ensure that every iteration brings real value. They also act as a communication tool, making it easier to explain progress to clients. Instead of just presenting numbers, you can show clear comparisons that highlight improvements. Here you can find a notebook with a full implementation from this blog post. The post How To Build a Benchmark for Your Models appeared first on Towards Data Science.

At long last, you can play as General Grievous in Fortnite, so long as you have the Chapter 6 Galactic Battle Pass. Grievous and his second outfit, Warlord Grievous, are locked to it, but unlike previous special pass skins there aren’t any specific quests you have to complete to unlock him. Below we explain how to unlock the General Grievous skin in Fortnite. How to get the General Grievous and Warlord Grievous skins in Fortnite All you have to do to grab both Grievous skins is own the battle pass and level up. You need to gain 28 new levels to unlock both versions of the skin, but if you just want the base General Grievous look, then you only need 14 levels. You can see the Warlord Grievous skin below. Between the two skins, there are also cosmetics that unlock every two levels, like a “Magnaguard Electrostaff” pickaxe and a “Separatist Cape” back bling. A few of the cosmetics can actually be claimed even if you don’t have the battle pass unlocked, as sweet little freebies. The full list of Grievous cosmetics and the levels required to get them are as follows: “Tsmeu-6 Wheel Bike” glider (two levels) “Grievous Grasp” emote (four levels) “Grievous vs. Windu” loading screen (six levels, no battle pass required) Magnaguard Electrostaff” pickaxe (eight levels) Grievous banner icon (10 levels) “Separatist Cape” back bling (12 levels, no battle pass required) General Grievous skin and “Dual Staff Slam” built-in emote (14 levels) “Saber Collector” wrap (16 levels) “The Soulful One” guitar/back bling (18 levels) “Warlord’s Origin” loading screen (20 levels, no battle pass required) “Kaleesh Cape” back bling (22 levels) “Grievous Stare” emote (24 levels) “Soulless One Dropper” contrail (26 levels, no battle pass required) Warlord Grievous skin (28 levels) In short, if you’re a battle pass owner, you should be able to unlock Grievous just by playing the game.

8BitDo Arcade Controller (Xbox/PC) $100 | Releases July 25 Preorder at Amazon 8BitDo Arcade Controller (Switch/PC) $90 | Releases July 25 Preorder at Amazon 8BitDo has revealed a pair of exciting new controllers for fighting game enthusiasts. Launching July 15 for Switch/PC and Xbox/PC, the 8BitDo Arcade Controller is a thin, leverless fight pad. All-button arcade "sticks" have become quite popular in the fighting game scene in recent years; some players find that the dedicated directional and jump buttons make it easier to hit precise inputs for special moves and combos. However, leverless models tend to be expensive, especially if you're looking for a reliable and customizable device. At $90-$100, 8BitDo's Arcade Controller will be the most affordable option from a major brand, and both models offer wireless connections on their respective platforms as well as PC. You can preorder the 8BitDo Arcade Controller for Switch/PC and Xbox/PC at Amazon. 8BitDo Arcade Controller (Xbox/PC) $100 | Releases July 25 As mentioned above, the new leverless 8BitDo Arcade Controllers use a button-only layout. Both models include four directional buttons (up, down, left, right), eight face buttons, and four additional remappable buttons. Each button uses hot-swappable Kailh Wizard low-profile mechanical switches, and the extra buttons also come with swappable locking flat caps in case you don't want to use them and want to avoid accidental presses. Both models also feature platform-specific inputs--like the Xbox Guide button and Switch Home button--for navigating the console's menus and interface.These leverless models are thinner and lighter than the previously released 8BitDo Arcade Controllers with joysticks, with a smudge-resistant material on the top panel, and a slip-resistant silicone mat on the bottom so it stays in place while playing. They also have foldable magnetic feet so you can adjust the angle.The Xbox version of the leverless 8BitDo Arcade Controller connects wirelessly to Xbox Series X|S, Xbox One, or PC with the included 2.4GHz dongle. The included USB-C cable can be used to establish a wired connection as well as charge the Arcade Controller.While the hot-swappable PCBs are convenient for modding the device's buttons and switches to your preferences, you can also personalize the button mapping and other options on PC or Xbox using 8BitDo's Ultimate Software X. One of those customizations is the RGB lighting rings around the buttons, which are exclusive to the Xbox model. Preorder at Amazon 8BitDo Arcade Controller (Switch/PC) $90 | Releases July 25 The Switch version of the leverless 8BitDo Arcade Controller the same button layout and overall design as the Xbox edition, but it's $10 cheaper and adds Bluetooth support. Since it uses Bluetooth to sync wirelessly with your Nintendo Switch, you should have no problem pairing it to the upcoming Switch 2 console.The Switch edition still comes with a 2.4GHz dongle for wireless play on PC, just like the Xbox model. The detachable charging cable can be used for wired mode on Switch or PC. So if you're buying the controller for PC, you can save $10 by opting for the Switch version. You will miss out on the backlit buttons, but the Switch edition does have a cool retro-themed design.The gray, black, and red color scheme pays homage to the Nintendo Entertainment System. The trio of bright menu buttons running along the top nod back to the SNES controller's face buttons.Button mapping and other competition-focused control customizations can be applied to the Switch edition using 8BitDo's Ultimate Software V2. Preorder at Amazon Gallery For those who prefer the traditional feel of a joystick, the original 8BitDo Arcade Controller is also available now. Like the levelerless edition, the 8BitDo Arcade Controller with a joystick is available in Switch and Xbox versions, and both work with PC. The Switch/PC model is $90, while the Xbox/PC model is $120 and available in either black or white color options. Continue Reading at GameSpot

EXCLUSIVE: Disney Lorcana's next set, Reign of Jafar, will launch in a matter of weeks, and we've got a first look at the Enchanted version of Captain Hook, The Pirate KingTech14:00, 16 May 2025Updated 14:36, 16 May 2025In case you missed it, Disney and Ravensburger's Lorcana is likely to go from strength to strength in its upcoming sets.Melding an easy-to-learn, hard-to-master card game with a huge slate of recognisable characters that should entice collectors as well as players, Lorcana has grown rapidly.‌After all, it's not easy to compete with Magic: The Gathering or the Pokemon TCG, but from our preview of the upcoming sets, it's clear Disney Lorcana is going nowhere.‌The game's next set, Reign of Jafar, hits prerelease on May 30 and lands everywhere else on June 6, meaning we're just a few weeks away from not just new cards, but 'Enchanted' versions of them, too–and Ravensburger has given us the opportunity to exclusively reveal one.Captain Hook appears in two different versions in the new setCaptain Hook will arrive in two guises in Reign of Jafar. Captain Hook, Forceful Duelist is a Common, while Captain Hook, The Pirate King is a Rare.Article continues belowIt's the latter that you can see above, with the Peter Pan baddie appearing in a slick new foil variant as part of the Enchanted lineup for the new set.He's seen with a sandy ship, presumably conjured by Jafar as part of the set's story which sees the Vizier shaping the world to his whims.If you've not seen the card before, Captain Hook, The Pirate King costs 5 Ink to play, and can be used in your Inkwell.‌The base art of the cardHe's a four-strength, five-willpower card and has two Lore. His ability, 'Give 'Em All You Got' says:"Once during your turn, whenever an opposing character is damaged, your Pirate characters get +2 Strength and gain Resist +2 this turn. (Damage dealt to them is reduced by 2.)"Article continues belowAs for type, he falls into Floodborn, Villain, King, Pirate and Captain roles.The artwork for this variant is by Nicola Saviori, with Kenneth Anderson being responsible for the standard printing of the car.For the latest breaking news and stories from across the globe from the Daily Star, sign up for our newsletters.‌‌‌