Nvidia spent a whopping $20 billion late last year to license Groq’s AI accelerator tech and hire away key members of its engineering team in an everything-but-the-kitchen-sink deal. The acquihire technically left Groq’s core inference-as-a-service business intact, but was clearly architected in such a way as to fly under regulators' radar. Only it didn’t. This week, The New York Times reported that the US Department of Justice had launched an antitrust probe into the deal. It’s hard to argue that Nvidia didn’t strip the startup for parts. It may not have been a merger in the traditional sense, but without its engineering staff, Groq may as well be Nvidia’s puppet at this point. Despite this, Nvidia contends the deal is a great American success story. “The Groq story is a prime example of the American system working as designed to promote innovation, reward entrepreneurs, and benefit consumers. The law is designed to encourage America's startup ecosystem and promote the fundamental rights of inventors and workers to pursue their dreams,” an Nvidia statement provided to El Reg and other media reads. Whether the acquihire of Groq actually harmed competition is another matter entirely. But, even if the Justice Department did force Nvidia to unwind the team, it’s probably too late. What exactly did Nvidia buy? Nvidia’s Groq acquihire bought it two key assets: mature silicon and the talent necessary to continue its development. Groq – which, by the way, is completely unrelated to Elon Musk’s Grok model series – made a name for itself using SRAM-heavy dataflow accelerators to speed up LLM inference to hundreds and now thousands of tokens a second, something that GPU-based systems from Nvidia had struggled to do on their own. But while faster than GPUs, the accelerators couldn’t achieve rapid throughput. Think of it this way: If Groq’s LPUs were the F1 cars, Nvidia’s GPUs were more like a city bus. But combine the two and you get something more akin to a sport pickup. At GTC in March, Nvidia unveiled its LPX racks, which are powered by 256 Groq-3 accelerators. As we understand it, they are really lightly modified versions of the startup’s existing Groq-2 chip designs, which makes sense, because three months is absurdly fast to tape out new silicon. Nvidia CEO Jensen Huang promised Groq-3 combined with its Vera Rubin GPU racks would deliver optimal performance across the entire spectrum of inference workloads. But, as we’ve discussed at length now, disaggregated compute architectures are not unique to Groq. Nvidia rival Cerebras is building similar systems with AWS and AMD, SambaNova is working with Intel, and d-Matrix and its partners are combining its in-memory compute platform with Nvidia GPUs to the same end as Nvidia’s Vera Rubin-LPX rack combo. The damage, if any, has been done Deals of this sort that are engineered to avoid regulatory scrutiny should get it anyway, several US senators have argued. While Nvidia didn’t outright buy Groq on paper, it may as well have. However, the real question for the DOJ is whether the deal was harmful to competition, and given the competitive landscape, proving harm may be easier said than done. But even if the DOJ found reason to litigate and was successful in unwinding the deal — it certainly wouldn’t be the first time regulators had torpedoed an Nvidia deal — it probably wouldn’t change much. Before Nvidia and Groq announced their licensing deal, the GPU giant was already laying the foundations for an ecosystem with its networking business at its center. In late 2024, the company contributed its MGX rack designs to the Open Compute Project (OCP) making it possible for any chipmaker to put their chips in racks originally designed for Nvidia GPUs. Then in mid-2025, GPUzilla opened its high-speed interconnect tech — the secret sauce that makes six dozen GPUs behave as one — to the broader industry through a licensing scheme called NVLink Fusion. As we recently discussed, the combination of open racks and Nvidia networking effectively meant that any chipmaker licensing the tech could slot their designs directly into Nvidia’s racks. If the DOJ blocked the acquihire and unwound the deal, Nvidia might lose direct control of LPU development and any revenues from the sale of the chips, but it wouldn’t necessarily be the end of its Groq LPX racks. Groq would just join the growing number of Nvidia hardware partners designing around the company’s AI factory ecosystem. In fact, if anything, the acquihire ensures that even if regulators eventually derail the deal, there will be plenty of alternatives lined up and ready to fill the void. ®
An autonomous machine gun system engaged drone targets in the Arizona desert, according to official photos from a U.S. military exercise. The system, called Bullfrog and built by Texas startup Allen Control Systems, pairs a standard M240 machine gun with a robotic turret, in this case bolted into the bed of a pickup truck. The […]
While high memory costs have hurt PC shipments, the server market continues to grow as AI infrastructure spending spreads beyond hyperscalers to corporate and government buyers. According to market intelligence firm IDC, the second quarter was a bumper one for the server sector, with vendor revenue reaching an all-time high of $166.3 billion. That was a 52 percent increase from the same period last year. The picture for servers therefore differs from that for laptops and desktops. There, unit shipments have fallen as buyers are discouraged by higher prices, driven by shortages of memory components. Yet higher prices have helped larger vendors sustain their revenue. In contrast, server shipments increased by 15.4 percent year-on-year in Q2, despite average selling prices being pushed up by elevated memory pricing and continued supply issues with other components. IDC said average selling prices increased across both GPU-accelerated and non-accelerated systems. Average selling prices for GPU-accelerated servers rose by nearly 44 percent to $170,200, even as GPU unit shipments fell 10.8 percent year-on-year. For non-accelerated systems, average pricing was up by more than 33 percent to nearly $13,000. AI infrastructure investment from hyperscalers and large cloud providers remains the largest source of demand, IDC observes. GPU-accelerated servers for the AI market made up nearly 53 percent of total revenue during Q2. However, it also says that AI server adoption is broadening beyond the largest players into enterprise and government-directed deployments across a growing number of countries, a policy and capex-driven layer of demand that is largely insulated from near-term commercial budget cycles. "The notable shift in the server market this quarter is in who is now buying," said Kuba Stolarski, IDC research vice president for Computing Platforms and Service Provider Infrastructure. "Demand is broadening beyond the largest hyperscalers toward specialized cloud providers (or neoclouds), sovereign AI programs backed by public capital, and enterprises beginning to adopt agentic and inferencing workloads," he added. Non-x86 servers now account for 44.8 percent of all server market revenue, according to IDC. That share has fallen from the first quarter, when they made up nearly half the total, despite the actual revenue figure rising from $58.7 billion to $74.4 billion. Another trend highlighted by IDC is that the big brands are starting to eat into the share of original design manufacturers (ODMs), the so-called white box server makers that have traditionally met the requirements of the hyperscalers. While ODMs collectively still make up the lion's share of server market revenue, this fell from over 60 percent last year to 53.9 percent in Q2. Leading the way is Dell Technologies, whose share rose from 7.7 percent a year ago to 13.4 percent. Supermicro is the second largest player, with 6.1 percent, followed by Lenovo on 5.1 percent, while HPE came fourth with 3.5 percent. The United States remains the biggest server market, generating $112.2 billion in Q2, or 67.4 percent of global revenue. China generated $26.4 billion, while Asia-Pacific excluding China and Japan reached $10.9 billion. Western Europe generated $9.1 billion and Central and Eastern Europe $0.7 billion. ®
AI infrastructure startup d-Matrix on Thursday joined the growing list of chipmakers licensing Nvidia’s NVLink Fusion interconnect tech and rack-scale reference designs to make its high-performance inference platform more accessible to customers. Under the deal, d-Matrix will integrate support for NVLink Fusion, a high-speed chip-to-chip interconnect that Nvidia began licensing last year, into future chip designs, including its upcoming Raptor accelerators. As we’ve previously reported, by embracing the tech, d-Matrix sidesteps many of the challenges associated with scaling its chip architecture across large compute clusters. By using NVLink over alternative interconnects and designing its compute blades around the GPU giant’s MGX reference designs, its customers can deploy its chips using the same racks and NVSwitch fabrics as Nvidia. By the end of next year, d-Matrix expects to offer systems with up to 144 Raptor accelerators connected by a single all-to-all NVLink fabric. While there’s a lot we don’t know about Raptor just yet, at the Hot Chips conference last month the company revealed each Raptor “card” would feature 32 GB of ultra-fast 3D-stacked DRAM on board capable of delivering 100 TB/s of memory bandwidth — roughly 4.5 times the memory bandwidth of Nvidia’s Rubin GPU. By the looks of things, the XPUs that will power d-Matrix's NVL144 racks are about half the size of the Raptor cards shown off at Hot Chips and include around 16 GB of 3D-DRAM and about 50 TB/s of memory bandwidth — still quite respectable by any measure. With 144 of these per rack, d-Matrix is looking at about 2.3 TB of memory capacity — enough for models exceeding four trillion parameters in size at 4-bit precision — and about 7.2 petabytes a second of peak aggregate memory bandwidth. This is achieved by bonding compute logic atop a stack of DRAM. The result is an in-memory compute platform that offers modest capacity while maintaining memory bandwidth closer to that of SRAM than is achievable using HBM. Memory bandwidth, as you may recall, is the biggest bottleneck for AI inference. The faster your memory, the faster the system can spew out tokens. This is exactly why Nvidia dropped $20 billion last year to license Groq’s IP and hire away its engineering talent. The chip’s SRAM-heavy dataflow architecture was capable of hitting 150 TB/s per chip, but the tradeoff is that SRAM isn’t very space-efficient and the chipmaker could only pack 500 MB of it onto a single die. d-Matrix's Raptor promises to deliver a decent fraction of that bandwidth with 64x higher capacity, which means the company can get away with using far fewer chips per model. Where a trillion-parameter model might need more than 2,000 Groq 3 LPUs at 8-bit precision, a single d-Matrix system would only need about 64 (32 at 4-bit precision). Just like Groq’s LPUs, d-Matrix chips can be deployed standalone, or as part of a heterogeneous compute config using GPUs for the compute-intensive prompt processing (prefill) phase of the inference pipeline and its Raptor accelerators for the memory bandwidth bound token generation (decode) phase. And for enterprises interested in the ultra-low latency inference capabilities of Groq, d-Matrix’s chips may offer a cheaper point of entry since fewer XPUs would be required. Nvidia’s walled garden We looked at Nvidia’s emerging IP licensing strategy in more detail last week, but in a nutshell, Nvidia stands to gain a lot more than licensing revenues from NVLink Fusion adopters. In addition to licensing Nvidia’s interconnect tech, d-Matrix plans to pair its accelerators with the GPU giant’s Vera CPUs, NVSwitch appliances, BlueField and ConnectX NICs, and SpectrumX Ethernet products. In other words, Nvidia stands to make a lot of money even if it’s not selling GPUs. It seems a fair number of chip designers are willing to make that kind of deal to avoid having to design their scale up networks or rack systems. Last week, MediaTek joined Marvell, Qualcomm, Arm, Fujitsu, and Amazon Web Services in adopting Nvidia’s NVLink Fusion interconnects. Nvidia is so invested in getting folks into its walled garden that it's spending billions on incentives to get them in the door. As part of the MediaTek deal last week, Nvidia invested $3.5 billion in the SoC designer, while it spent $2 billion to get Marvell in the door. No word on whether the latest deal included any such terms. ®
The U.S. Navy has added $38.2 million to a BAE Systems contract to overhaul two Mk 45 Mod 4 naval guns, the main deck gun mounted on the Navy’s destroyers and cruisers. Naval Sea Systems Command issued the cost-plus-incentive-fee modification to an existing contract, funded through fiscal 2024 shipbuilding money that will remain available beyond […]
UPDATED OpenAI will design its next-generation AI accelerators with the help of South Korean foundry giant Samsung Electronics. This move could help the ChatGPT maker with memory, process tech, or other key things it needs to succeed. The announcement, made by OpenAI Korea general manager Harrison Kim in a press conference Wednesday, comes just weeks after the company showed its first-generation “Jalapeño” accelerators besting Nvidia’s Blackwell GPUs across a slew of inference workloads. "One of the areas where we have made the most progress and gained the most recognition with Samsung Electronics is our joint production and research on the next-generation chips we are developing," he said, according to Reuters. Following publication, OpenAI got in touch to clarify that Kim's comments were made in reference to OpenAI's existing relationship with Samsung announced last year, not a new deal. "The reference to expanding our collaboration with Samsung was made in the context of our existing relationship, including Samsung Electronics’ adoption of ChatGPT and the LOI announced last October. We have nothing new to announce. We did not announce plans to manufacture chips through Samsung," a OpenAI spokesperson told El Reg after the story published. Now that the hype has died down and Wall Street has convinced themselves that the rise of AI ASICs isn’t the threat to Nvidia’s dominance they feared, OpenAI still has to contend with the realities of being a fabless chip designer. The vast majority of leading edge silicon made today comes from a single company: TSMC. That means securing meaningful capacity requires competing with AMD, Nvidia, and practically every other chip biz and hyperscaler designing their own chips. But TSMC isn’t the only leading edge wafer fab out there. They’re just the biggest. Smaller, less popular options include Samsung Electronics and Intel Foundry. The supply of high-bandwidth memory (HBM), which OpenAI’s (and Broadcom if we’re being honest here) designs make extensive use of, is similarly constrained with just three major manufacturers capable of producing it in volume: SK Hynix, Micron, and you guessed it, Samsung Electronics. This puts OpenAI in a tough spot of having to compete with larger, higher volume chip designers for both logic and memory silicon. As such, OpenAI’s tie-up with Samsung could go two ways. The most obvious, but also the least interesting, is that Samsung will supply all or at least some of the HBM used in the model dev’s next-gen chips, which we’ll call Habanero for lack of an official code name. Detailed at Hot Chips last month, OpenAI’s Jalapeño AI inference chip uses HBM4, the same tech used by AMD and Nvidia’s latest GPUs. That tells us that there's a pretty good chance Habanero will likely use the faster HBM4E. And given the ongoing memory shortage, if OpenAI wants to secure ample supply, it’s going to need to forge these relationships early. Besides supply chain, timing is important because there’s a good chance OpenAI’s next chip will take advantage of custom base dies to move things like memory controllers or even some logic from the compute silicon to the HBM stack itself. If you recall Nvidia’s disclosures around NVHBM a few weeks back, it’s the same idea. However, these custom base dies require some coordination with the HBM supplier, so fostering a close relationship with Samsung now would benefit OpenAI, even if the compute chiplets themselves end up being manufactured by TSMC. While the most obvious reason for OpenAI’s tie-up with Samsung is memory-related, it could conceivably fab its next chip on Samsung’s 2 nm process tech just to have an alternative supplier to TSMC. Samsung is the second largest leading edge foundry operator in the world and used to produce a lot of silicon for the likes of Nvidia, Apple and others. Nvidia’s Ampere generation used Samsung’s 8 nm process tech for the majority of its PCIe-based cards, while its HBM-equipped accelerators were exclusively fabbed by TSMC. But a number of factors, including the rise of multi-die architectures requiring advanced packaging and broader adoption of HBM-based accelerators, made TSMC the more attractive option. Reported challenges with Samsung’s yields across its 3 nm process tech probably didn’t help either. But as of 2026, Samsung is looking a lot more compelling, with chip designers including Rebellions, Tenstorrent, and Groq building sophisticated AI accelerators at its fabs. Rebellions’ Rebel100 accelerator is a particularly useful proof point as it utilizes a multi-die architecture and HBM3E, both of which require sophisticated advanced techniques to stitch everything together. The chip shows that Samsung is more than capable of producing OpenAI’s next chip. What’s more, doing so could open the door to larger allocations than they would get from TSMC alone. ® Updated 9/10 at 2100 GMT to reflect that this is not a new deal.
Nearly every country represented in the United Nations General Assembly recently voted in favor of promoting world maps that more accurately show the size of various continents. But the UN resolution is unlikely to change how millions of people regularly see versions of 16th-century Mercator maps depicting skewed continent sizes in navigation apps.
On September 4, the African-led resolution attracted 164 votes in favor, with six countries abstaining from voting. Only the United States voted to oppose the resolution after describing it as a “radical ideological project,” according to UN News.
The non-binding UN resolution does not impose map changes on anyone. Instead, it “encourages governments, schools, international organizations, and technology companies to use the Equal Earth projection and other so-called equal-area maps when relative size matters,” UN News reported. The resolution also encourages teaching about the trade-offs that arise from depicting a spherical planet on a flat map.
Elbit Systems has introduced FUSE, a new business line bringing together the Israeli defense contractor’s autonomous drone and ground vehicle systems under one portfolio, the company announced September 9 in Rosh HaAyin, Israel. FUSE centers on software called Dominion-X, which lets a single operator control multiple autonomous aircraft and ground vehicles at once, a setup […]
China’s military is exploring humanoid robots for urban combat as the country’s commercial robotics industry dominates global shipments and manufacturing.
China’s military is exploring humanoid robots for urban combat as the country’s commercial robotics industry dominates global shipments and manufacturing.
Startup Bluecore Energy has raised $50 million in seed funding to develop floating nuclear power plants intended to supply ports and other locations with rapidly growing energy demands. The California-based company said the Silverton Partners-led round, which includes $10 million in previously announced pre-seed funding, brought in several new investors. It plans to use the money to engineer and test its compact nuclear power system and pursue regulatory approval and maritime classification. Its initial design is intended to generate about 10 MWe – megawatts of electrical output – continuously using light-water reactor technology. Bluecore says the maritime system would be able to operate for years at a time and require refueling only once every few years. The company has secured its first barge and is developing a non-fueled, electrically heated reactor module prototype at its Port of Long Beach headquarters to test monitoring, sensor, and control systems. Bluecore is targeting ports, AI datacenters, and coastal infrastructure, with the reactors potentially deployed offshore and connected to customers by subsea cables. Remote islands and coastal settlements in places such as Alaska may also lack practical connections to a wider electricity grid. Bluecore claims its approach could provide reliable, zero-emission power to energy-intensive locations and open additional markets for nuclear generation. "Six months ago, we were building the foundation. Today, we have the capital, the team, the hardware, and some of the most important institutions in nuclear and maritime working alongside us," said Bluecore founder and CEO Kofi Asante. "Our focus now is simple: create and deliver zero-emission energy as safely and quickly as possible." The move follows the launch of the International Atomic Energy Agency's (IAEA) Atomic Technologies Licensed for Applications at Sea (ATLAS) initiative in the US capital last month. Secretary of Energy Chris Wright linked that event to the Department of Energy's own efforts to rapidly expand the provisioning of atomic energy generating capacity in order to meet the requirements of AI datacenters and other industries. Bluecore has begun discussions with the US Nuclear Regulatory Commission (NRC) and Coast Guard. The two agencies recently signed a memorandum of understanding to coordinate oversight of the design, construction, and operation of civilian floating nuclear power plants. The NRC says it is ready to license maritime reactors under existing federal regulations, while it develops an additional framework for microreactors and similarly low-risk designs. Bluecore remains years from commercial deployment. It has yet to build a fueled reactor or apply for the approvals required to operate one, and its Long Beach proposal must also navigate California's restrictions on new nuclear power plants. While some may have concerns about waterborne atomic power plants, America has a long history of maritime nuclear generation. All of the US Navy's active aircraft carriers are nuclear-powered, but the technology first went to sea aboard the submarine USS Nautilus in 1955. USS Enterprise became the first nuclear-powered aircraft carrier in 1961, while the cruiser USS Long Beach, commissioned the same year, was the first nuclear-powered surface combatant. Bluecore says its team includes former US Navy nuclear submarine officers and leaders with experience across organizations including SpaceX, Northrop Grumman, Toyota, Rivian, and Uber. Russia currently operates the world's only floating nuclear power plant. The barge-like Akademik Lomonosov is considerably larger than Bluecore's proposed design and can generate up to 70 MWe. ®
ASML and chipmaking giant TSMC have launched an industry-wide push toward 12-inch photomasks, saying the larger format could lower costs and overcome some limitations of high-NA EUV lithography. Dutch firm ASML, the only commercial supplier of EUV lithography systems, says it and TSMC have established a collaboration intended to lead the transition to larger-format photomasks for extreme ultraviolet (EUV) lithography. Intel Foundry and Samsung Electronics have also voiced their support. The initiative aims to establish a 12-inch mask pilot line by 2031 and have the supporting lithography systems ready for advanced node production by 2033. According to ASML, high-NA EUV will initially enter production using current 6-inch masks, but moving to 12-inch masks could increase fab productivity, lower costs, and remove stitching constraints. The push for larger masks stems from the anamorphic optical design ASML adopted for its first generation of high-NA EUV machines so chipmakers could continue using existing 6-inch masks. The "NA" in high-NA EUV stands for numerical aperture – a measure of an optical system's ability to collect and focus light. Achieving a higher NA required larger optics, which increased the angle at which light strikes the reticle, or mask, creating shadowing and contrast problems. ASML says it could have used optics that reduced the mask pattern by 8x in both directions, rather than the 4x used in existing systems, but that would have required larger masks. Instead, it adopted anamorphic optics that reduce the image by 4x in one axis and 8x in the other, allowing chipmakers to continue using 6-inch masks. The trade-off is an exposure field half the size of that in previous machines. Large dies that exceed this field may therefore require two separately exposed patterns to be stitched together, adding complexity and reducing productivity. ASML and TSMC now see larger masks as a way to restore the full exposure field as the industry pushes toward denser chips and smaller process nodes. "We expect the adoption of High NA EUV to increase progressively along the device scaling roadmap, first using current 6-inch masks and then further supported by 12-inch masks, which enable greater scanner productivity and allow the industry to meet the demand for smaller, faster and more energy-efficient chips," said ASML president and CEO Christophe Fouquet. TSMC has so far not used high-NA EUV, including for its 2 nm process, but says it intends to deploy the technology in high-volume manufacturing for advanced nodes beginning in 2030. It expects the number of layers requiring high-NA EUV will rise as new process nodes are introduced, driven by increasingly complex transistor architectures required for AI applications. Intel Foundry signaled its support for the initiative, with EVP Naga Chandrasekaran saying: "Within our lithography capabilities, Intel Foundry is focused on near term enablement of high-NA on 6-inch masks with or without stitching, and making the transition to 6x12-inch masks. We will continue to work closely with ASML and the entire industry to enable this transition." Samsung is also backing the initiative and says it plans to become the first chipmaker to introduce ASML's high-NA EUV technology into high-volume DRAM manufacturing by 2028. "The AI era is transforming the semiconductor industry and increasing the importance of technological innovation across the entire value chain," said Samsung Electronics vice chairman and CEO Young Hyun Jun. "By further strengthening our collaboration with ASML, we are helping lay the foundation for the next generation of AI and semiconductor innovation." IDC senior research director Andrew Buss told The Register that high-NA EUV creates design and manufacturing challenges for future products. Although reticle stitching and chiplet designs can mitigate these, he said, the move to 6 x 12-inch photomasks is a necessary step comparable to the industry's transition from 200 mm to 300 mm wafers. Buss said the transition would require toolmakers, chip manufacturers, mask suppliers, and designers to align their technologies and processes, making broad industry participation essential. ®
Sometimes you might run the same model twice and get different results. That often happens when you’ve upgraded it with skills. Skills are detailed text documents that lay out the tools the model should use, the approach it should take and how it should analyze the results. Good skills are practical, pulled from actual reports on HackerOne and other bug bounty platforms. A model can still lean on its own knowledge, but that’s just less efficient.
There are plenty of skills out there you might come across, but not everything can be trusted. Some skills can simply be dangerous and infect your system. To make sure they are safe, you can check them with SkillSpector by NVIDIA, so you don’t end up with anything malicious on your system.
Bug Bounty Skills
Both of these repositories do bug bounty hunting end to end, but they go about it in almost opposite ways.
The first is called Bountyforge. It’s actually just one single skill file, but it’s smart enough to split itself into eight different mini agents that all work at the same time. One looks at websites and apps, another at crypto and blockchain, others go after different angles hackers can exploit. It also checks each finding with four different tests to make sure it’s not a false alarm. Then you get a report in whatever format the bug bounty program wants.
You don’t even need Claude Code or any other coding tool for this, you can just run it right inside the regular Claude website in your browser.
The second bug bounty repository is Claude-BugHunter. It takes the opposite approach. The repo has 83 skills and almost half of those were built by studying 681 real bug reports that people actually got paid for on HackerOne. These skills aren’t locked to Claude Code either, you can use OpenCode, Codex or Hermes Agents with them.
Here are a few examples of the results we got with these skills.
API endpoints are often vulnerable and this is worth trying your luck on to see how it goes.
Another approach can be APK reverse engineering. Here we found a hardcoded RSA-2048 signing private key baked into the published APK. With that key, hackers can push a new app to the app store and infect every employee phone, getting access not just to the WiFi network at the workplace but to their personal life too. Quite dangerous.
We found an API endpoint vulnerable to an SQL injection and managed to pull the entire database.
Having skills built on real attacks keeps the model from wandering off into its own weird approaches and missing a lot of good findings.
Active Directory Skills
Claude-AD was made by ADScanPro for testing a company’s internal network. It gives your model a playbook with skills and agents built for an Active Directory assessment. The developers are upfront that it’s not an auto pwn tool. It’s meant to guide you through the assessment. Every finding can get mapped to a compliance control (DORA, NIS2 and ENS).
Claude-AD is very careful about getting caught too. It explains what a security team would actually see on their end if that technique got used. And any time it’s about to do something that would actually change things on the company’s network, it stops and asks for confirmation first.
General Cybersecurity Skills
Antropic-Cybersecurity-Skills is basically a giant reference book. It has 817 skills covering 29 areas of security work, cloud security, malware analysis, all the way down to hardware and firmware. Each skill is its own small file, so your agent will quickly pull out the two or three it actually needs for its task.
Every skill ties back to real security frameworks that companies and auditors already use (NIST CSF, MITRE ATT&CK and so on). So if your model finds a problem using one of these skills, it can also tell you exactly which official standard it violates. You can use it to justify findings to a compliance team.
SCADA Skills
On an industrial network, a clumsy scan can shut down a production line or damage physical equipment, since a lot of this gear is old and wasn’t built to handle unexpected traffic. That’s why the ICS skill by Masriyan is built to never actively touch a live industrial network. Instead, it works off network captures someone already took. It reads the file, recognizes industrial protocols by the ports they normally run on (Modbus, DNP3, Siemens S7, EtherNet/IP, OPC-UA, and more) and counts which devices are talking to each other. It then shows you write commands, these are the ones that change a value on an industrial device. That’s the traffic you want to see first.
The second mode skips network captures and instead searches for exposed industrial equipment using Shodan and Censys. The skill can also help your model reason about how an industrial network is laid out and check findings against MITRE’s ICS specific attack framework and the IEC 62443 security standard.
Science Skills
Although science isn’t really what we want to focus on here, in one of our SCADA articles we mentioned that to carry out a successful attack requires hackers to understand the technical process of the plant. That means understanding how the chemicals are produced and which units are used along the way. We also showed how vinyl acetate is produced and talked about paracetamol production.
1 kg of paracetamol at 100% purity was reported to cost €8,205, while 1 kg at 99% purity cost just €5. So even a single day of sabotage could cause serious financial damage to an enterprise.
Finding a scientist among hackers is quite a challenge, which is why Stuxnet needed a group of people from different backgrounds working toward one objective. But now hackers can just import different skills to make their attacks more devastating. K-Dense published 140 skills with access to different scientific databases and Python tools.
The real concern here isn’t ICS exploits inside the repository, there aren’t any. It’s the access to sensitive scientific data paired with an AI agent that can actually understand that data and change it.
Summary
AI skills can be a gamechanger, especially when they’re based on actual reports hackers got paid for. These skills show your model how to approach things and what tools to use during the test, so it doesn’t wander off hallucinating and inventing its own ways of testing things. That can wreck your bug bounty flow, since you’ll end up overlooking plenty of potential targets.
Simply relying on the AI to find things isn’t enough, hunters that do it keep getting a lot of dupes. You need to test things manually too. For this reason we created our Bug Bounty training to show you how to find bugs and work with the AI more efficiently.
Arm's latest smartphone platform is optimized for running AI agents on mobile devices and introduces neural graphics that it claims can deliver desktop-class gaming within a smartphone's power constraints. Announced at the Arm Everywhere China event in Shanghai today, Compute Subsystem (CSS) for Mobile 2 succeeds last year's Lumex CSS and reflects Arm's embrace of AI as a strategy for future growth. Arm sees itself as central to this shift, arguing that AI workloads currently run largely in the cloud but will increasingly move to devices at the network edge. "When every interaction has to travel to a datacenter and back, there's implications for latency, connectivity, the cost of delivering those services at enormous scale, so this is where we're seeing it becoming more distributed, and I think few doubt now that basically everything you can push to the edge you will," says Arm's EVP of Edge AI, Chris Bergey. AI has evolved from background capabilities such as speech recognition and image processing into tools people actively engage with, including generative assistants, Bergey claims. Agentic AI goes further by interpreting objectives, coordinating applications and models, and executing sequences of actions. "So you have repeated inference, application logic, system orchestration, multiple workloads operating concurrently, and so that means you don't just need more CPU performance. You need multiple high-performance CPUs working in parallel to keep all those different workloads running," Bergey explains. With this in mind, the C2 CPU cluster features a new top-end C2-Ultra core that Arm claims delivers up to 15 percent higher single-thread performance than last year's C1-Ultra. The C2-Ultra cores have a larger execution engine to keep more instructions in flight, and Arm claims more accurate branch prediction and better fetch and target prediction to keep the instruction pipelines fed. Arm's chipmaker licensees can mix and match components to suit their requirements, as usual, but the firm touts a flagship cluster with two C2-Ultra cores, six efficiency-focused C2-Pro cores, and two SME2 units (Scalable Matrix Extensions), compared with one SME2 unit in last year's platform. Doubling the SME2 capability enables up to a 70 percent speedup on the latest small language models, according to Arm, while the cluster topology supports private caches on the individual CPUs plus a large shared L3 cache. All of this is tied together by the DSU, or the DynamIQ Shared Unit. Arm says the cluster has been optimized to reduce latency and keep concurrent agentic workloads responsive. CSS for Mobile 2 also introduces dedicated neural acceleration for graphics in the Mali G2-Ultra NX, which Arm calls its first AI-native Mali GPU. Arm previewed the neural technology last year, giving developers time to prepare for the hardware. The GPU introduces an all-new execution engine, neural accelerator hardware, and a third-generation ray tracing unit, built for full-scene lighting and shadows. The hardware supports Neural Super Sampling (NSS), which upscales game graphics from 540p to 1080p, and Neural Frame Rate Upscaling (NFRU), which generates intermediate frames to increase the frame rate. Using both techniques means the GPU renders one eighth of the pixels contained in the final sequence of displayed frames, with the neural hardware reconstructing the rest. Arm claims this makes desktop-class graphics possible within a 1 W power budget. "The G2-Ultra NX GPU is able to sustain a 30 frame per second gaming session on content with very rich and intricate ray traced scenery. And this is really enabled thanks to a 70 percent reduction in the overall ray tracing workload, which then translates to a 30 percent boost in performance," says GPU Product Manager Deyan Lazarov. Arm also demonstrated other games running at 60 frames per second. Like the CPU cores, Arm is offering Premium and Pro variants of the Mali G2 GPU, allowing its chipmaker customers to trade off performance against cost. To encourage uptake, Arm has published its neural graphics SDK and sample code on GitHub, with support for the cross-platform Vulkan graphics API. As with previous Arm smartphone platforms, chipmakers must now turn the designs into working silicon before smartphone vendors can bring devices to market, possibly as early as next year. ®
Login
