LesenForschungRadarAnlageframework
Anmelden / Registrieren
AlleTechnikProdukteInvestierenBörseSonstiges

Lesen

2026-11-1246 Beiträge

Aktualisieren

Quadric CMO Says RISC-V for AI Acceleration Is a Bad Idea; Breker CEO Argues Control Is the Correct Positioning

Wesentlich
2026-11-12 16:01 GMT+8

Steve Roddy, chief marketing officer at AI chip company Quadric, made a sharp judgment in a Semiconductor Engineering roundtable discussion: using RISC-V for AI acceleration is a "completely bad idea." His argument is not based on RISC-V's inherent lack of capability, but on the cascading consequences of architectural choices—once you create a partition between a programmable processor and an external accelerator, you simultaneously create three problems: data movement, compiler complexity, and interconnect verification.

Roddy's mechanistic reasoning is specific. CPUs excel at pointer chasing and unpredictable thread switching—with 40 threads running, you have no idea what will execute in the next clock cycle. But AI inference does not follow this workload pattern: you know what needs to be done millions of cycles in advance. Therefore, most CPU hardware logic is useless for inference. More critically, while DSPs are closer to what is needed, they are vector machines (vector-by-scalar or vector-by-vector), whereas the most challenging operations in AI are matrix-by-matrix. Thus, his conclusion is that if you start from a CPU to build an accelerator, you are essentially using the wrong starting point to solve a problem that shouldn't exist, then spending billions of dollars to make up for that initial error.

However, Dave Kelf, CEO of verification company Breker, provided counter-evidence from actual projects. He observed that large AI accelerators—whether SIMD or SIMT architectures—are traditionally controlled by complex state machines, which he describes as "hopeless." Four companies he has worked with (both large and small) are replacing state machines with RISC-V processors to drive their accelerators. The reasons are practical: RISC-V allows adding custom instructions for efficient load/store operations to reduce partition latency, and integrating vector or matrix multiplication instructions to handle residual computation when a parallel lane in a SIMT array diverges.

Kelf cited a specific case: a customer had a SIMT array where occasional divergence occurred in some parallel lanes during massive parallel math operations. When divergence happened, they used the RISC-V control processor to handle the residual computation for that lane, while the rest of the channels continued running on the accelerator. This was not heavy inference on the RISC-V—it was just "enough math processing to keep the whole machine running." He acknowledged this required significant engineering effort, but the results were "quite successful."

The consensus of the roundtable discussion thus became clear: RISC-V’s value in AI systems is not to perform inference calculations itself, but to serve as an orchestration layer—using custom instructions to manage data flow between accelerators, reduce partition latency, and handle edge cases. Roddy also conceded that "you can get value out of tweaking and tuning a processor for the control function," though he did not believe it should be used for heavy inference.

But the discussion revealed a frequently overlooked layer: correct instructions do not guarantee a working chip. Ewald Liess of SignatureIP pointed out that when multiple RISC-V cores coexist with an accelerator in the same system, cache-coherency corner cases and traffic bursts can lead to interconnect deadlocks. An instruction may execute perfectly at the ISA level, but if the required data burst is unavailable, the chip will still fail. Francisca Tan of Siemens EDA added the dimension of observability—tracing custom instructions, system-level performance monitoring, and embedded analytics are indispensable during silicon debugging.

Quellen:semiengineering.com

Forschung

US cybersecurity firm Cloudflare researches MOLE post-quantum anonymous credential mechanism for AI agent traffic management

Kurzfassung
2026-11-12 08:00 GMT+8

US-based cybersecurity and cloud computing company Cloudflare recently disclosed that it is researching a new post-quantum secure anonymous credential mechanism called MOLE (Moderation Of unLinkable Endorsements). The research aims to address the threat of future quantum computers with cryptographic attack capabilities potentially breaking traditional encryption technologies, which could allow attackers to forge tokens and bypass security controls such as rate limiting.

The existing Privacy Pass technology allows users to prove they have passed verification while protecting their privacy, but websites can still identify the source of the token issuance. Cloudflare's MOLE attempts to further hide the user's endorsement source: the verification service can confirm that the endorsement comes from a trusted institution without identifying exactly which party provided the endorsement, subsequently issuing a credential for the user to access the website.

Technically, Cloudflare is researching the combination of multivariate cryptography and VOLE-in-the-Head (VOLEitH) zero-knowledge proof technologies to build an anti-quantum anonymous credential system suitable for MOLE. Beyond basic security protection, Cloudflare has also listed AI agent traffic management as a key future application direction for MOLE, hoping to use it to reduce the situation where massive automated AI requests occupy website resources and affect the experience of ordinary users.

However, the technology is currently in the research and development stage. Preliminary tests show that although some solutions can complete a single operation within milliseconds, post-quantum cryptographic technologies generate larger signature and verification data, with some operations requiring the transmission of several KB of data. If applied to actual web services in the future, besides verification speed, the communication burden caused by frequent verification will need to be carefully considered.

Quellen:iThome

Forschung

Susannah Streeter, Chief Investment Strategist at Wealth Club: Huang’s Scandal Exposes Key Person Risk for Nvidia, Valuation Multiples Face Compression

Wesentlich
2026-11-12 19:35 GMT+8

Nvidia's stock price has recently come under pressure, but Susannah Streeter, Chief Investment Strategist at Wealth Club, believes that what truly warrants investor attention is not the short-term decline, but rather the pricing logic of the valuation premium being re-examined. Speaking on the podcast Zero Sum, she pointed out that Nvidia's position as a core supplier of AI infrastructure means a significant portion of its valuation premium is bound to CEO Jensen Huang personally—he helped investors understand the scale of the AI opportunity and positioned Nvidia as the central computing infrastructure supporting this wave of construction. When questions arise regarding Huang's judgment, this binding relationship creates a disproportionate impact.

Streeter's core mechanism is that even if chip demand remains robust, a shock to leadership confidence could trigger compression of valuation multiples. Investors will require the company to demonstrate a broader leadership bench and clearer governance structures, while regulators and lawmakers are tightening scrutiny around AI and chip exports. This means the premium is no longer supported solely by revenue growth and market share; it also requires governance credibility as a prerequisite. She cited Apple's transition from Steve Jobs to Tim Cook as an example, illustrating that founder-driven companies can enter new phases, but they need different types of leaders to guide the company to the next stage.

The second-order effects of this incident extend to US policy on chip exports to China. Streeter argues that the scandal has added a national security dimension to an already highly politicized debate. She expects lawmakers to seek greater oversight of advanced chip exports, including measures requiring companies to track the overseas flow of chips to prevent smuggling. Washington may also pressure allies to restrict exports of semiconductor manufacturing equipment to China. With Democrats controlling both the House of Representatives and the Senate, she anticipates increased pressure for AI regulation and more restrictions on the influence of large technology companies.

For investors, Streeter's advice is not to immediately sell off Nvidia shares, but to reassess portfolio concentration. She warns that investors may have exposure to a select group of US technology companies through global funds that exceeds their expectations, and they should ensure adequate diversification across industries and geographies. She simultaneously distinguishes between short-term volatility and the long-term prospects of the AI industry: Nvidia remains a key supplier with strong demand, but not every company associated with the AI boom will become a long-term winner—she draws an analogy to the dot-com bubble period, where many star companies failed while a few, such as Amazon, became true beneficiaries.

The applicability boundaries of Streeter's argument lie in the speed of the transition and the degree of regulatory escalation. Following her analytical logic, if Nvidia's board quickly completes a credible leadership transition and export policies do not tighten further, the company's profitability and premium may be re-established. Furthermore, she maintains a cautious stance on infrastructure investment—companies investing huge sums without proving sufficient returns face valuation pressure, whereas underlying infrastructure suppliers such as power, grid, and utilities will still receive support even if data center construction slows. Her overall judgment is that the investment narrative for the AI sector is shifting from simple demand growth to a comprehensive assessment of governance quality, regulatory environment, and capital return rates.

Quellen:invezz.com

Forschung

"AI-Driven Data Foundation Design" Author Sanjeev Mohan: Agents Bypass Pipelines to Build Their Own Databases, Observability Must Be Embedded in Architecture

Thema · AI智能体数据访问模式Wesentlich
2026-11-12 19:02 GMT+8

For decades, enterprise data architecture followed a fixed path: data was generated once in business systems, then pushed through pipeline stages such as extraction, transformation, cataloging, and preparation before being consumed by humans for batch reports. Sanjeev Mohan believes the premise of this model is now being broken. His assessment is that when thousands of agents run simultaneously, they will not wait for pipelines to finish—they demand direct access to the database.

Mohan describes a behavioral pattern never before seen in software: agents create short-lived database instances, copy data within them, parallelize tests across different reasoning chains, and only then select a course of action. He cited a loan approval example: an agent might query census records and social media data simultaneously before rejecting an application. This immediately raises a governance question—how is that decision audited?

In his view, this differs fundamentally from past data governance. Past governance was deterministic: manage the data, then run a fixed reporting dashboard on top of it. Agent governance faces constantly changing context at runtime. For the same task, the context could be entirely different depending on the person (or the specific agent instance). Therefore, he stated that observability and transparency must become intrinsic parts of the architecture, rather than things bolted onto a data warehouse afterward.

From this, Mohan derives a further conclusion: future platforms will need to fuse structured and unstructured data into a single converged system, and new types of databases will emerge—specifically designed to store runtime state and build runtime context. He has not yet provided the specific form or implementation path for such databases; this judgment remains a proposal regarding architectural direction.

It should be noted that all of the above are personal arguments articulated by Mohan during an interview with theCUBE at KubeCon North America. As the head of SanjMo Consulting and author of "Designing the AI-Driven Data Foundations," he has long studied data architecture issues. However, the text provides no measured data from deployed systems or independent verification. The value of this argument lies in its precise naming of a constraint: agents' data consumption patterns differ from those of human consumers, meaning pipeline architectures designed around human consumption will constitute bottlenecks.

Quellen:SiliconANGLE

Forschung

VulnCheck Report: Only 217 of Nearly 40,000 Candidate Vulnerabilities Discovered by Anthropic's Glasswing Are Publicly Fixed

Wesentlich
2026-11-12 15:55 GMT+8

Cybersecurity firm VulnCheck released a report in November 2026 independently tracking Anthropic's "Project Glasswing," an initiative designed to use Claude models to discover security vulnerabilities in open-source software at scale. The report indicates that seven months after the program's launch, Claude had identified 37,966 candidate vulnerabilities, yet only 217 were marked as fixed in public records as of early November.

This data contrasts with figures disclosed by Anthropic itself. According to official data from November 2, out of 37,966 candidate findings, 6,376 had been reviewed by external experts, 5,087 were reported to maintainers, and 629 received upstream patches. While differing statistical methodologies account for some discrepancies in the number of fixes, both parties acknowledge a core issue: the speed of AI-driven vulnerability discovery has created a severe bottleneck in human review capacity, significantly limiting disclosure and remediation efficiency.

To alleviate this pressure, Anthropic launched the OSS Scanner service on October 8, allowing open-source maintainers to receive model-generated vulnerability reports directly, bypassing Anthropic's internal human review process. However, this "zero human intervention" model shifts the validation burden to the community. For instance, Daniel Stenberg, maintainer of the curl tool, recently received 45 vulnerability reports from Anthropic. He stated that the core maintenance team was overwhelmed by the influx of reports and had to change rules to discuss low and medium severity vulnerabilities publicly on GitHub, leveraging community help for review.

VulnCheck's report notes that this gap is accelerating. In just the last two months, candidate vulnerabilities surged by nearly 12,000, while the record of fixed issues increased by only 15. Additionally, the 442 CVEs attributed to Anthropic tracked by VulnCheck represent only 0.5% of the global public CVE total for 2026. Although Anthropic did identify serious flaws like a Use-After-Free memory error in curl via Claude, the overall picture suggests that AI's powerful discovery capabilities have not translated into proportional remediation outcomes, instead imposing a heavy review workload on open-source maintainers.

Quellen:量子位

Forschung

Liao Hung-yuan, Director of the Institute of Information Science, Academia Sinica: Large Models Can Describe Images but Fail to Understand Human Actions; Modular Visual Understanding Is Needed to Bridge the Gap

Thema · 廖弘源模块化视觉理解Wesentlich
2026-11-12 18:04 GMT+8

When a photo of a child is input into a large vision-language model, the model’s description is merely "Everything is normal." Liao Hung-yuan, Director of the Institute of Information Science at Taiwan's Academia Sinica, uses this example to illustrate that while current general-purpose vision models possess considerable image-description capabilities, they struggle with more nuanced judgments closer to real-world needs—such as what the person is doing, whether danger is occurring, or what relationships exist between objects in the environment.

Liao Hung-yuan believes that to bridge this gap, Academia Sinica’s direction is not to build another large general-purpose model, but to develop specialized visual-understanding capabilities that complement large language models, including behavior analysis, object understanding, and edge computing. His reasoning is straightforward: edge devices have limited computing power and memory, making it impossible to run large models directly on cameras. Therefore, multimodal AI must adopt a modular approach, allowing different models to handle tasks suited to their strengths.

Home care is the most intuitive validation scenario for this architecture. When a camera detects someone falling, the system cannot simply know "there is a person in the frame"; it must determine if the person has fallen and requires assistance. If help is deemed necessary, the system uses text-to-speech to ask, "Do you need help?" After the user responds, automatic speech recognition interprets the voice content, which is then connected to a large language model. Finally, the system notifies family members or emergency services while simultaneously relaying the result back to the person who fell.

The core of behavior analysis is Taiwan GBA, developed by Academia Sinica in collaboration with Chung Yuan Christian University and National Tainan University. Its key concept is the Universal Action Space (UAS), inspired by DeepMind’s Kinetics-600 dataset—which collected videos of 600 types of human actions, each lasting about 5 to 10 seconds. The research team used these videos to establish an action feature space, enabling the model to learn correlations between different actions. In the future, new behaviors such as falling, waving, or attacking do not require building completely independent models; instead, they can be mapped into the existing action space. Liao Hung-yuan describes UAS as a "comprehensive sports stadium": different sports or behaviors can all be understood within this space.

However, human behavior cannot be judged from a single image; it requires understanding temporal changes. Two photos showing a person standing versus lying on the ground may appear as different states, but what truly matters is the process "from standing to falling." The research team utilizes Video Swin Transformer to analyze dynamic features in videos, identifying locations in the images related to actions—for instance, when playing tennis, the model focuses on the person’s hands, racket, and body movement areas—and converts these dynamic details into trainable behavior features, mapping them into the universal action space.

Quellen:iThome

Forschung

Google and Flight Attendants Union Reach Settlement to Scrub Spirit Airlines Data

Thema · 精神航空数据脱敏协议Wesentlich
2026-11-12 18:15 GMT+8

According to WIRED, Google and the Association for Flight Attendants, the union representing crew members of the defunct Spirit Airlines, reached a groundbreaking settlement this week. The agreement addresses privacy concerns arising from Spirit Airlines' plan to sell its internal data assets to Google for AI model training during its liquidation process. Previously, Spirit sought to sell decades of data, including billions of emails and messages, for $10 million, but the union argued that these records contained significant personal employee information.

Under the revised terms, Spirit Airlines must hire a third-party data de-identification firm called Tonic AI to remove information potentially linked to individuals or consumers and ensure the anonymized data cannot be relinked to specific people. Google has committed to maintaining the data only in this newly de-identified form. The agreement explicitly excludes highly sensitive records, such as employee medical files, documents related to investigations and labor relations, and information about industry-wide safety management systems. A US aviation trade organization also intervened, warning that including such data could discourage employees from reporting safety issues through voluntary programs, which are critical to the industry's safety record.

The settlement includes an opt-out clause allowing Google to withdraw from the sale if the data-cleaning process reduces the asset value by more than 8 percent. Although viewed as a strong protection for employee privacy, the deal still requires judicial approval expected next week. Sara Nelson, president of the Association for Flight Attendants, stated that this outcome could serve as a template for other unions, which plan to demand similar data protection clauses in future labor agreements. The union is also working with state-level politicians to legislate worker data protections. This event highlights the growing tension between labor rights and the commercialization of private data amid the AI industry's surging demand for training materials.

Quellen:WIRED

Forschung

AMD Zen Original Architect Mike Clark Explains Zen 6 Design Shift: Dense Cores Become Highest IPC Performance Core

Thema · AMD Zen 6架构Wesentlich
2026-11-12 06:22 GMT+8

Mike Clark, AMD's original Zen architect, explained a fundamental design shift during an AMD In-Depth interview with Chips and Cheese: starting with the Zen 6 generation, AMD's dense technology cores are no longer area-optimized products but serve as the highest IPC performance cores. He explicitly stated that while dense technology was previously just an area trade-off, this generation's team leveraged it to add extra IPC within the core. The classic design is now repositioned as the high-frequency version, while dense technology serves as the top-tier performance core because full L3 cache has been enabled in this implementation, providing 4MB per core, totaling 128MB for 32 cores, and delivering the highest achievable IPC.

The mechanism behind this reversal is not mysterious. Clark explained that because Zen 6c targets lower frequencies than the high-frequency cores, it can accommodate more capacity in its queues—the floating-point scheduler is just one example; other queues throughout the core are larger in Zen 6c, aiming to extract more IPC and gain leverage in SMT environments. He simultaneously emphasized a design principle: there should be no software-differentiating delta between the two architectures, meaning the same optimizations apply to both.

Regarding AI inference capabilities, Zen 6 adds two VNNI pipelines on the vector side. Clark admitted this was not the initial design goal—they did not use the term 'agentic' when they began designing. However, Zen 5 had already established a new foundation for AI workloads, and the team always believed the core had a place in AI: orchestration, data type manipulation, and small model inference. Observing the direction of inference workloads, they decided to add VNNI pipelines. Since these are long-cycle projects and the floor plan was already set, this was a relatively late decision, but the team ultimately succeeded in squeezing in the extra pipelines.

There is a counter-intuitive reduction in the execution engine: the number of integer multipliers decreased from three in Zen 5 to two. Clark explained that Zen 5 used a more unified scheduler, where adding a third multiplier to the scheduling group was simple and did not increase complexity. However, Zen 6 required returning to a more distributed scheduler to scale the architecture; that extra multiplier did not yield significant gains in Zen 5, so to optimize area and the scheduler, they reverted to two multipliers.

The frontend also saw improvements. Zen 5's dual-decode setup made it difficult for a single thread to synchronize instruction cache and operation cache fetches in larger workloads. Zen 6 spent considerable time optimizing this, improving efficiency from both delivery and power perspectives. Regarding branch prediction, misprediction recovery latency was reduced, and accuracy saw a slight improvement. Clark stated they do not intend to expand the branch prediction structure but rather layout TAGE tables more efficiently, while focusing each generation on restarting execution as quickly as possible after a misprediction.

Quellen:chipsandcheese.com

Forschung

South Korean Humanoid Robot Startup Robros Announces World Action Model, Targets 2027 Open-Source Release

Thema · Robros世界动作模型Kurzfassung
2026-11-12 14:18 GMT+8

South Korean humanoid robot startup Robros announced its latest technology roadmap on November 10 during the 'NVIDIA Physical AI Startup Connect' session at NVIDIA AI Day Seoul 2026. The presentation was led by Im Dae-gyu (임대규), the company's Chief AI Officer (CAIO). Key announcements included a 'Real-to-Sim-to-Real' training methodology and a proprietary 'World Action Model' (WAM).

According to the launch details, the 'Real-to-Sim-to-Real' approach involves creating a digital twin of actual work environments within a virtual space. Robots are trained in this simulation using NVIDIA's Isaac Sim, Isaac Lab, and Omniverse platforms, before the learned actions are transferred to physical robots for validation. This method aims to reduce the time and cost associated with repeated trial-and-error in the real world and mitigate dangerous operations. Robros has already reconstructed its laboratory environment using this method and validated box and manufacturing part handling tasks on its proprietary humanoid robot, IGRIS-C.

Additionally, Robros revealed plans to develop the WAM, an AI model designed to understand surrounding environments and generate corresponding robot actions. Currently, the model is undergoing pre-training using simulation data, videos of human operations, and data collected from actual robot deployments. The company explicitly stated that it targets releasing the WAM as an open-source project in 2027. If realized, this would provide global robot developers with a foundational action-generation model based on a simulation-first strategy.

Robros indicated that it will expand proof-of-concept (PoC) initiatives in manufacturing and logistics while simultaneously advancing its own AI model development and field data accumulation. Im emphasized that deploying humanoid robots in industrial settings requires data reflecting diverse working environments and validation technologies, stating that the company will enhance robot operational capabilities through its proprietary platform and simulation-based training.

Quellen:스타트업 생태계의 모든 것 \'플래텀(Platum)\'

Forschung

Google Cloud Releases 2027 Security Trends: Proposes 'Beyond Zero' Architecture to Manage AI Agent Risks

Wesentlich
2026-11-12 08:00 GMT+8

U.S. cloud provider Google Cloud released its 2027 enterprise security trend forecast on November 12, 2026, stating that artificial intelligence will further reshape the cybersecurity landscape. On the defensive side, AI agents are expected to assist Security Operations Centers (SOC) in automatically investigating and handling incidents; however, on the offensive side, software supply chain contamination, data extortion, and abuse of AI models and cloud computing resources will become primary threats.

To address risks arising from excessive permissions granted to AI agents, Google Cloud proposed the "Beyond Zero" security architecture. This approach argues that enterprises cannot merely verify whether a user or AI agent has permission to access an application; they must also restrict accessible resources and executable operations based on security policies and real-time risk. This shift aims to prevent attackers from using techniques like prompt injection to induce high-permission AI agents to execute malicious actions and move laterally across other enterprise resources.

The report also highlighted management blind spots caused by "Shadow AI Agents." As employees or departments build AI agents without integrating them into IT security management, these agents may unknowingly access sensitive data or consume cloud resources. Google Cloud recommends limiting tool usage permissions for such agents and tracking their compute usage and costs. Additionally, Mandiant, a subsidiary of Google, disclosed "LLMjacking" attack cases, showing that attackers use stolen API keys to impersonate enterprises and call large model services, with some victims accumulating up to $1 million in unauthorized fees.

Quellen:iThome

Forschung

South Korean Tech Giant Naver Signs ~2 Trillion KRW AI Compute Deal, Plans Independent AI Agent App and Logistics Investment

Thema · Naver AI战略Wesentlich
2026-11-12 14:19 GMT+8

South Korean internet giant Naver disclosed major strategic developments during its third-quarter earnings conference call on November 12, 2026. CEO Choi Soo-yeon stated that the company has signed a computing service supply contract worth approximately 2 trillion KRW (about $1.4 billion) with a US-based AI inference company. The deal covers about 20 megawatts (MW) of computing power over four to five years, with initial revenue expected to begin in 2027. This order marks the first large-scale external customer for Naver’s “AI Factory” business, validating its strategy to convert internal infrastructure capabilities into B2B services.

The contract includes Nvidia B300 chips and the next-generation Vera Rubin platform. Naver emphasized that it has proactively secured most domestic data center capacities equipped with high-density power and liquid cooling facilities, noting that contracted capacity already exceeds its 2028 target of 200 MW. CFO Kim Hee-chul revealed that negotiations are underway with at least ten other potential clients, with further announcements expected. Naver previously reduced capital expenditure pressure through an investment agreement with Brookfield Asset Management and participation in Nvidia’s equity increase, adopting a model where an asset-holding company purchases GPUs while Naver leads operations.

Beyond infrastructure, Naver plans to launch an independent AI agent application by the end of 2026. This app will not only integrate generative AI features within search but also serve as a standalone service capable of performing actual tasks for users, such as bookings and payments. To adapt to the AI agent era, Naver’s advertising billing model will shift from “pay-per-click” to “value-based” charging tied to actual transaction outcomes, with a pilot program scheduled for the first half of next year.

In the e-commerce sector, Naver announced the establishment of a dedicated logistics subsidiary, planning to invest nearly 1 trillion KRW by 2030 to enhance delivery competitiveness. Data shows that free shipping and return policies increased member purchase frequency by 27%, and Q3 N Delivery transaction volume grew 38% year-on-year. By building its own logistics infrastructure and using shopping data to predict demand, Naver aims to shorten delivery times and optimize inventory distribution, consolidating its advantage against competitors like Coupang.

Quellen:바이라인네트워크

Forschung

Unit 42: Indicators for AI Autonomous Penetration Tools Fail Quickly; Behavioral Features Offer More Persistent Detection Value

Thema · ARTEX AI渗透框架Wesentlich
2026-11-12 07:00 GMT+8

In a report released on November 11, Palo Alto Networks threat research team Unit 42 presented a judgment with direct operational significance for defenders: when facing AI autonomous penetration tools, indicator-based detection is failing rapidly. The ARTEX framework is frequently forked and modified, and attackers rotate infrastructure at an extremely fast pace, leaving almost no defensive window before traditional indicators such as IP addresses and user agent strings lose their value. However, based on observations from multiple intrusion investigations and managed threat hunting, Unit 42 points out that there is a class of signals more durable than indicators—namely, behavior itself.

This judgment is grounded in a structural fact: autonomous penetration frameworks rely on the foundational reasoning patterns of large language models to drive decision-making. Regardless of how code is forked or infrastructure rotated, an agent that waits for model inference to complete before initiating the next round of requests will inevitably exhibit specific network behaviors. Typical patterns observed by Unit 42 include: bursts of one to fourteen requests sent at approximately one-second intervals, followed by pauses of twenty-five to thirty-one seconds, consistent with the time an agent spends waiting for a model inference call to complete; lateral movement completed within about one minute after obtaining credentials, with verification of over two hundred client secrets within three minutes; immediate extraction of routing targets from downloaded OpenAPI specifications or source code maps; and failed requests being adjusted and retried successfully within seconds. While these characteristics appearing individually might merely indicate tool configuration anomalies or manual testing, their combination points to autonomous agents.

Based on this, Unit 42’s defensive recommendation is to build detection around operator behavior rather than solely tracking indicators. Specifically, defenders should look for canary usernames (such as placeholder values like evil[.]com or test@example[.]com) in identity provider logs, monitor the first use of cloud keys from new IP addresses, identify write-then-delete test patterns on storage buckets, and focus on early probing stages preceding later intrusions. The report simultaneously acknowledges a limitation: in at least one observed case, the attacker switched from scripted clients to a desktop browser for the most sensitive steps, suggesting human intervention or browser automation may break pure behavioral patterns. Therefore, behavioral detection is necessary but not sufficient, and must still be combined with authentication hardening, attack surface reduction, and default-deny strategies.

Quellen:Unit 42

Forschung
Weiter →
Anmelden / Registrieren
LesenForschungRadarAnlageframework

The practical implication for chip architects from this discussion is that RISC-V’s customization capabilities have clear applicability boundaries in the AI context. It is suitable for orchestration, control, and handling edge cases; it is not suitable for being a matrix inference engine. But even within its "suitable" range, between correct instructions and a usable system lie three verification hurdles: cache coherency, interconnect deadlocks, and performance observability.

In home care applications, the research team tested using Nanyang Technological University’s NTU RGB+D dataset, which contains 220 types of human actions. The team selected nine care-related actions for analysis. Liao Hung-yuan stated that in the dataset testing environment, the model’s recognition rate reaches approximately 80% to 90%. However, he also acknowledged that the biggest challenge in actual deployment is not laboratory data, but real-world data. Many crime datasets consist of simulated scenarios; for example, someone holding a knife might be acting, differing significantly from a real crime scene. For AI to be truly applied on-site, more real-world data accumulation is still required.

Regarding environmental understanding, Academia Sinica researcher Dr. Wang Chien-yao extended the YOLO v7 and YOLO v9 architectures to develop G-YOLO (Generalist YOLO). Traditional YOLO excels at real-time object detection, quickly answering "what things are in the picture," such as identifying people or cars in a street view. But for self-driving cars or robots, what is truly needed is understanding relationships between objects: if a group of people is crossing the road, AI must not only know "there are people" but also understand that this group is crossing the road, identify nearby buildings, and comprehend the relationships between people and surrounding objects. With G-YOLO added, the model can generate more complete scene descriptions, such as noting that approximately 15 people are crossing the road, with shops, buildings, and road facilities nearby.

Currently, Taiwan GBA can recognize 12 behaviors, including falling, fighting, and brandishing knives, covering application scenarios in home care, public safety, and campus security. G-YOLO continues to expand depth estimation and multi-view understanding capabilities, enabling AI to know not just "what was seen," but also "where objects are, how far away they are, and how they interact with each other."

Liao Hung-yuan believes that in the future, self-driving cars, robots, and smart care systems will require not only large language models but also a set of multimodal capabilities that can understand the real world. Large models allow AI to begin describing the world, but to truly enter the world, AI still needs to understand human actions, environmental changes, and the meaning behind events. This is the core direction of Academia Sinica’s investment in autonomous multimodal AI research and development.

Vector loads increased from two in Zen 5 to four (256-bit). Since Zen 5 could already perform two 512-bit loads, the 1,024 wires were already present; splitting their usage was simply more complex than using them contiguously. The L2 cache bandwidth issue was also resolved: Zen 5 prefetchers would sometimes preempt L2 bandwidth, whereas Zen 6 ensures prefetches and demand requests can run in parallel continuously, delivering the full 64 bytes per cycle.

Interviewer George Cozma reported a measured result: Zen 6c matched AMD's fastest Zen 5 server CPU in SPECint 2026, despite the frequency disadvantage. This provides a concrete anchor for Clark's IPC argument, but remains a vendor-side observation based on a single benchmark test and has not yet undergone independent verification.

All technical details stem from first-party statements by AMD architects. Clark also acknowledged that some content constitutes undisclosed secrets—for instance, the specific techniques for maintaining bandwidth within the L3 mesh. Readers should view these as explanations of design intent rather than independently verified performance guarantees.