LectureRechercheRadarCadre d'investissement
Connexion / Inscription
ToutTechnologieProduitsInvestissementBourseAutres

Lecture

2026-11-1392 publications

Actualiser

Nvidia Unveils Scale-In Architecture Using BlueField-4 DPUs for Hardware-Level Agentic AI Security

Matériel
2026-11-13 06:46 GMT+8

Nvidia Corp. recently unveiled "Scale-In," a new pillar in its AI networking infrastructure designed to address the security and management challenges of agentic AI in data centers. As AI agents assume increasingly autonomous roles within enterprise environments, traditional security boundaries face significant strain. Scale-In integrates BlueField-4 Data Processing Units (DPUs), the DOCA software platform, and Spectrum-X Ethernet to deliver purpose-built acceleration and security for AI factories.

Gilad Shainer, Senior Vice President of Networking at Nvidia, explained in an interview with SiliconANGLE that agentic AI presents a completely different use case: "Now we’re dealing with many agents that need to run on the infrastructure... You need to increase security because you need to make sure that those agents are contained in their sandbox and are not doing the wrong things." The core of Scale-In lies in hardware-enabled "out-of-band" monitoring. BlueField-4, Nvidia's fourth-generation DPU, features in-silicon security engines that collect telemetry from memory, storage, and networks independently of the GPU and CPU. This design ensures that the security monitoring layer remains isolated from the agents themselves, preventing them from tampering with or influencing their own audit trails.

Beyond hardware isolation, Nvidia has deployed OpenShell, an open-source runtime tool designed to govern agent execution, access, and inference routing. Combined with BlueField-4, OpenShell creates a comprehensive security stack ranging from software-defined boundaries to hardware-enforced controls. Shainer noted that while OpenShell manages agent behavior, BlueField-4 provides a "completely separated, completely isolated" device to monitor the environment, ensuring no agent can influence the monitoring process. While elements of this approach appeared in previous reference architectures like the integration of Vera Rubin infrastructure and Vera CPUs, Scale-In formally establishes this as a distinct pillar of Nvidia's AI networking strategy.

Sources :SiliconANGLE

Recherche

Terminal-Bench 4.1 Uses Low-Privilege Execution for Isolated Scoring Logic to Prevent Agents from Tampering with Results

Matériel
2026-11-10 07:13 GMT+8

A benchmark score is a claim about capability, and the verifier is what makes that claim checkable. If the verifier itself can be manipulated by the object being evaluated, the score no longer measures capability but rather the model's ability to find scoring loopholes. The core change in Terminal-Bench 4.1 targets this structural risk.

Agent benchmarks differ from traditional static Q&A: models need to execute commands, write code, and modify files within a terminal. This means the artifacts submitted by the model may contain executable code, which could run inside the verifier container. If the submitted code can access the grader or its output files, it can interfere with the scoring process—not by earning rewards through task completion, but by directly modifying the reward values.

The solution adopted by Terminal-Bench 4.1 is RewardKit Workers: an optional package from the Harbor team whose core design treats code submitted by models as untrusted input. The Worker runs the submitted code as an unprivileged user, preventing it from reading test files or writing reward files, and terminates all processes it starts before returning. The verifier evaluates outputs independently in an isolated environment, keeping tests and expected answers on the scoring logic side. Structurally, the submitted code cannot access the scoring results it attempts to influence.

Why not use a simpler approach—patching known attack patterns one by one? The article points out that fixing individual known vulnerability patterns helps, but each added check might also reject legitimate work. Pattern matching is an adversarial approach, whereas structural isolation is a design approach: instead of guessing how a model will attack the scorer, make the scorer physically incapable of being attacked.

This mechanism has clear applicability boundaries. Version 4.1 uses Workers only in affected verifiers; not all 66 tasks have adopted this scheme. The Worker protects the grading reward, not a general-purpose machine sandbox—task-level sandboxing and resource limits remain important. Furthermore, even with isolation, task authors still need domain expertise to determine whether checks truly capture successful completion and allow legitimate alternative solutions.

A practical verification datum: re-scoring the 4.0 leaderboard trials using the 4.1 verifier flipped approximately 1% of the rewards, and leaderboard scores remained within previously reported confidence intervals. This indicates that 4.0 scores are largely valid but not entirely immune to verifier defects. For readers, this means that when interpreting any agent benchmark score, it is worth asking: did the verifier behind this score employ isolation?

Sources :Snorkel AI

Recherche

Google Launches Gemini 4 Hydrogen and Orchestrator Feature in Antigravity for Dynamic Model Switching

Matériel
2026-11-13 05:26 GMT+8

US tech giant Google launched the new language model Gemini 4 Hydrogen on November 12, 2026, along with a new feature called Orchestrator in its Antigravity development environment. This feature changes how coding agents use AI models: instead of keeping the most capable model running throughout an entire session, it dynamically switches between Hydrogen and the more powerful Gemini 4 Carbon as work progresses.

According to Google, this combination achieves nearly the same results as running Carbon alone while using 24% less quota and finishing tasks 19% faster. Under the Orchestrator mechanism, Carbon handles demanding reasoning tasks such as architecture design and debugging, while Hydrogen executes code changes and testing. For example, Carbon might analyze a repository to determine how to implement a feature, then hand off the actual coding to Hydrogen; if a test fails requiring a rethink, Carbon steps back in before passing execution back to Hydrogen.

Clement Farabet, VP of AI Engineering at Google DeepMind, described the combination as "a huge unlock for agentic coding" in a post on X. Varun Mohan, who leads Google's Antigravity team, noted that pairing Hydrogen with Orchestrator and swarms has become "very popular internally."

However, Google's own benchmark reporting contains inconsistencies: the main Gemini announcement cites Terminal-Bench 4.0, while the Antigravity announcement references version 4.1. Additionally, Google has not published Carbon's standalone success rate, per-task breakdowns, or sufficient details about Orchestrator's routing decisions to demonstrate where the reported efficiency gains originate.

Currently, the Orchestrator feature is limited to Google AI Ultra subscribers, with support for more models and subscription tiers planned. Hydrogen is rolling out to paid Antigravity users, with broader availability expected.

Sources :The New Stack

Recherche

Cloud Provider Nebius Claims Internal RL Auto-Tuner Achieves Up to 4.93x Training Speedup on Nvidia B300s

Prise rapide
2026-11-12 08:00 GMT+8

AI infrastructure company Nebius released a technical blog post on November 12, 2026, detailing an internally developed reinforcement learning (RL) auto-tuning system. According to the company, the system was tested on NVIDIA B300 Tensor Core GPUs across four models and two workloads, achieving training step speedups of 1.42 to 4.93 times on fixed GPU budgets, with a median speedup of approximately 3.5 times.

The tests covered dense and mixture-of-experts models including Qwen3.8-27B, Qwen3.6-35B-A3B, GLM-5.3-Flash, and DeepSeek-V4.1-Flash, running on math reasoning (DAPO) and agentic coding (SWE-bench Verified) workloads. The highest gain, nearly 5x, was observed for GLM-5.3-Flash on coding tasks. Nebius noted that DeepSeek-V4.1-Flash coding jobs ran faster on 40 GPUs than an early setup did on 56, reducing GPU-seconds per step by 38%.

The auto-tuning system works by profiling each job to identify bottlenecks, forecasting promising changes, and testing them in real runs. Nebius reported that its prediction subagent had a median step-time error of 5.7% in retrospective evaluation. All optimal configurations remained within the job's trainer-rollout gap limits, accompanied by numerical checks and reward comparisons to ensure training quality was preserved.

It is important to note that these figures are self-reported by Nebius under specific hardware conditions and model combinations, with no third-party independent replication or public benchmark code currently available. For developers relying on large-scale RL post-training, this efficiency gain could significantly reduce computational costs and time if reproducible in other environments.

Sources :nebius.com

Recherche

GitHub Developer Advocate Andrea Griffiths: A Three-Layer Separation Method for Building Agent Skills Portable Across Clients

Matériel
2026-11-13 06:54 GMT+8

How many times this week did you re-explain the same thing to the same agent? And how many times did you have to ask a colleague with deep institutional knowledge to solve a specific problem? In this practical guide published on November 12, 2026, Andrea Griffiths, a developer advocate at GitHub, identifies this repetitive labor as the core pain point that agent skills aim to solve: context should not disappear in chat logs, nor be locked inside one person's head.

An agent skill is a versioned directory centered on a SKILL.md file, describing when and how to complete a repeatable task. Griffiths emphasizes a key boundary: skills describe the instructions for doing things, not the permissions to do them. A skill can describe how to query a database, but it does not grant access to that database. Tools, credentials, and permissions are still provided by the host environment running the skill. This means the skill itself is portable, but portability does not equal plug-and-play—the target environment must already possess the tools and access conditions assumed by the skill.

With this boundary established, Griffiths presents the core method for making skills truly reusable across projects and teams: separating project configuration from skill instructions. In the Data Forge support-desk pipeline example, the forge-init skill first probes source tables and naming conventions, writing the discovered project details into a resolved manifest; subsequent skills read this manifest to obtain the project context they need, rather than guessing independently or hardcoding specific names into their instructions. The manifest describes the project name, metric store, and platform engine, while the skill describes the rules and standards for doing the work. A pattern found in a repository does not automatically equate to a rule that should be preserved, so discovered details must be reviewable, with parts requiring business judgment confirmed or modified by humans.

How multiple skills coordinate is the second key issue. The forge-builder skill coordinates the entire workflow from issue to pull request, calling forge-dbx to query the data platform and forge-lint to apply team standards. forge-dbx provides a unified path for data platform access for all Data Forge skills: it reads the manifest and selects either a Databricks or DuckDB adapter, so each skill does not need to define its own connection and query logic. Griffiths summarizes this design principle in one sentence: when steps change, you modify one skill, and the rest remain untouched.

Cross-client portability is the final link in this methodology. The Data Forge example retains skills//SKILL.md as the sole source file, using a packaging script to generate client-specific adaptation files: .claude-plugin/plugin.json for Claude Code, .cursor-plugin/plugin.json for Cursor, and .github/agents/.agent.md for GitHub Copilot. The script also includes a CI check mode to ensure client files remain synchronized with the source. After installing a skill, Griffiths recommends verifying item by item: whether the client recognizes the skill for the intended task, loads referenced files, can use the required tools, and returns the evidence requested by the skill.

Sources :The GitHub Blog

Recherche

Ariel Jirau, Head of Lab Engineering at Natera: When Diagnostic Accuracy Stalls, Check Data Coverage Before Tuning Models

Matériel
2026-11-13 05:47 GMT+8

Natera is a clinical laboratory company specializing in cell-free DNA (cfDNA) testing. Its lab operations rely on a wide array of precision instruments, each with distinct failure modes, vendor documentation, and resolution pathways. The service engineering team responsible for maintaining this infrastructure has long been organized around deep individual expertise rather than standardized operating procedures. When a senior technician departed, years of accumulated diagnostic knowledge were lost; no system existed to capture, organize, or scale-deploy these experiences. With over 65 tickets per day and an average first response time of 17.4 hours, instrument downtime directly delayed sample processing and patient reporting.

Ariel Jirau, Head of Lab Engineering at Natera, led the team in building the SWARM system in collaboration with AWS. In reviewing the project's architectural decisions, Jirau offered a judgment highly relevant to practitioners: diagnostic accuracy is fundamentally a data quality issue. The team found that when accuracy entered a plateau, the most common cause was not insufficient model capability but rather the absence of specific vendor manuals for certain instrument families in the knowledge base. After supplementing missing documentation for the lowest-scoring instrument family, accuracy rose from 71% to 75%, with no modifications made to the model itself.

This insight stems from a specific set of architectural choices. SWARM’s core component is a versioned skill library containing 33 skills: 24 instrument-specific skills covering all equipment families supported by Natera, and 9 cross-domain workflow skills applicable to all ticket types. Each skill has a defined interface, is independently testable, and undergoes review against the same standards as the overall system—it is treated as production software rather than embedded prompt logic. Jirau explained the rationale: when a diagnosis fails, the team needs to point to a specific skill to isolate the problem, rather than debugging the entire system.

Inside the skill library, there is a clear division of labor: Python handles deterministic tasks, such as parsing Jira fields or formatting HTML reports; large language model (LLM) skills handle linguistic reasoning, such as interpreting symptom descriptions, matching vendor documentation, and generating diagnostic narratives. This boundary is one of the most impactful decisions Jirau identifies when building multi-skill agents—it determines which processes can be precisely tested and audited, and which inevitably carry probabilistic uncertainty.

SWARM splits diagnostic work into two independent runtimes sharing the same skill library. The swarm-agent is designed for throughput: it selects skills based on equipment type and symptoms, queries the knowledge base, and generates a structured HTML report containing a primary cause hypothesis, ranked troubleshooting steps, and a self-assessed confidence label, delivering it to the ticket within minutes. The swarm-hive is designed for depth: engineers can engage in interactive follow-ups, upload images, and explore alternative troubleshooting paths, while AgentCore memory retains per-engineer context across sessions. This separation exists because autonomous background analysis and interactive engineering collaboration have fundamentally different latency and context requirements.

Sources :Amazon Web Services

Recherche

Google Releases Gemini 4 Hydrogen Model as New Default for AI Search

Matériel
2026-11-13 05:55 GMT+8

US tech giant Google announced on November 12, 2026, the release of its new Gemini 4 Hydrogen large language model, setting it as the default engine for its AI Mode search. Robby Stein, VP of Product at Google Search, described it as "the biggest model upgrade to Search yet." The model is available starting today across Search, the Gemini app, all tiers of Google AI Studio, the Gemini API, and Gemini for Business.

Google states that Gemini 4 Hydrogen is designed to provide more helpful answers to complex questions, including those with multiple parts, hard-to-understand typos, or highly specific requests. The new model can break down user questions and address each part in its responses. It also offers a wider range of links by "conducting simultaneous queries across all the relevant subtopics" and includes more visual elements like images, comparison tables, graphs, and maps. One example shared by Stein showed the AI generating an interactive animated driving simulation to explain how phones locate which lane of a highway a car is on in areas without cell service.

In terms of performance, Google claims Gemini 4 Hydrogen delivers "significant leaps in both speed and intelligence" compared to the previous Gemini Flash 3.8 model, and that it "matches or outperforms some of the competing models in its class," such as Anthropic's Claude Sonnet 5.5 and OpenAI's GPT-6.1 Terra. Users now have additional thinking level options, including low, medium, high, and a new "xhigh" mode. While these comparisons are based on Google's self-reported benchmarks and lack independent verification, the update marks the latest step in Google's deep integration of generative AI into its core search experience.

Sources :The Verge

Recherche

Russian Cloud Provider Yandex Cloud Reports 2.84% Permanent Data Loss After Drone Strikes, Mandates Default Backups

Prise rapide
2026-11-13 05:00 GMT+8

Russian cloud provider Yandex Cloud updated its customers via a blog post, stating that it had “restored access to more than 97% of customer data” following extensive damage to its data centers from Ukrainian drone strikes. The firm admitted that “the remaining 2.84% cannot be restored due to critical damage to equipment.”

Yandex attributed part of this unrecoverable loss to the lack of customer backups, noting that users who utilized Yandex Cloud’s backup tools “did not lose any data.” In response, the company has changed its policy so that “currently, all virtual machines created in Yandex Cloud work with the backup function enabled by default.” Additionally, Yandex opened a new availability zone, ru-central1-f, at an undisclosed location to improve fault tolerance.

Sources :Tom's Hardware

Recherche

Runtime Security Startup Raven.io Announces Entry into the Korean Market, Targeting Regulated Industries Like Finance

Prise rapide
2026-11-12 19:17 GMT+8

On November 12, 2026, Raven.io, a startup specializing in runtime application security, held its first domestic press conference in Seoul, officially announcing its entry into the Korean market. Founded by former Symantec executives, the company aims to help enterprises identify and block attacks on open-source components before Common Vulnerabilities and Exposures (CVE) numbers are released or patches become available, using its unique runtime monitoring technology.

Roi Abitboul, co-founder and CEO of Raven.io, emphasized during the presentation that as AI is widely used to discover and exploit vulnerabilities, traditional defenses based on known signatures are increasingly lagging. Raven.io's technology acts like a "closed-circuit television" installed on code execution paths, providing real-time visibility into running libraries, functions, and execution routes. When abnormal behavior is detected, the system immediately applies policies to stop dangerous actions without interrupting the entire application.

To establish a foothold in the Korean market, Raven.io has partnered with The Executive Action, a Singapore-based business consulting firm. The companies plan to enter the market through three strategies: a subscription model, reseller partnerships, and collaboration with Managed Security Service Providers (MSSPs). Given the strict requirements for data security and compliance in the financial sector, and its heavy reliance on complex SaaS and open-source infrastructure, Korean financial institutions will be their primary target customers.

Sources :바이라인네트워크

Recherche

CopilotKit Releases OpenBot 1.0: An MIT-Licensed, Self-Hostable AI Coworker Workspace with Built-In Governance

Matériel
2026-11-13 05:45 GMT+8

CopilotKit released OpenBot 1.0 on November 9, 2026, introducing an MIT-licensed, self-hostable AI coworker workspace. Unlike many vendor-locked AI agent products, OpenBot allows enterprises to utilize their own infrastructure, models, and agent frameworks while maintaining security through built-in governance mechanisms.

A central feature of the product is its "governance gateway." Every browser click, file write, or tool call passes through a policy evaluation based on Common Expression Language (CEL) before execution. Audit records are written prior to the action taking place, and broken or denying rules result in refusal rather than permission. This design addresses enterprise concerns regarding data leakage and unauthorized operations when deploying autonomous agents.

OpenBot supports a wide range of major AI agent frameworks and models. It is compatible with the AG-UI protocol, enabling integration with over eight frameworks including Claude Agent SDK, OpenAI Agents SDK, and LangChain. For models, users can employ OpenAI, Anthropic, Google Gemini, as well as open models like Kimi and DeepSeek. Each "Bot" runs in its own container with a dedicated Chromium browser and file system, accessing only granted tools.

For individual users, OpenBot offers a free local version; for enterprises, the business tier costs up to $29 per active user per month and includes Single Sign-On (SSO), SCIM synchronization, granular tool permissions, and full audit trails. The release also includes desktop applications for Mac and Windows to simplify local deployment.

Sources :MarkTechPost

Recherche

Applied Materials Reports Record Earnings: Q4 Revenue Up 53%, Non-GAAP EPS Doubles

Matériel
Vérifié 2026-11-13 06:39 GMT+8

US semiconductor equipment manufacturer Applied Materials, Inc. reported its financial results for the fourth quarter and fiscal year ended October 25, 2026, on November 12, 2026. The company achieved record quarterly revenue of $10.41 billion, an increase of 53 percent year over year. Non-GAAP earnings per share (EPS) reached $4.16, up 92 percent year over year.

For the full fiscal year 2026, total revenue was $34.45 billion, up 21 percent year over year. Full-year non-GAAP EPS was $12.89, an increase of 37 percent. Gross margins expanded significantly, with Q4 GAAP gross margin at 50.4 percent and non-GAAP gross margin at 50.6 percent, representing increases of 240 and 250 basis points respectively compared to the same period last year.

In terms of cash generation, the company produced record operating cash flow of $3.30 billion and free cash flow of $2.6 billion in the fourth quarter. It distributed $2.42 billion to shareholders through $2.00 billion in share repurchases and $421 million in dividends. The share buyback amount marked the largest quarterly repurchase in more than eight years.

CEO Gary Dickerson stated that AI continues to fuel strong incremental demand for semiconductors, positioning the company well for multi-year growth. CFO Brice Hill noted that technology leadership in critical areas of the wafer fab equipment market drove the significant year-over-year improvements in revenue and profitability.

Sources :sec.gov

Recherche

Sage CTO: Finance Teams Reject AI Not Because Models Are Inaccurate, But Because the Reasoning Process Is Invisible

Matériel
2026-11-12 18:06 GMT+8

Harris pointed out at the Sage Future for Partners conference that a year ago, the challenge for AI applications in finance was getting models accurate enough to be trusted. Now, the challenge is everything surrounding the model: the systems that enable agents to operate safely, auditable reasoning chains, and stable costs acceptable to CFOs. He breaks trust down into four specific conditions—data security, ethical construction, human control, and competent, predictable execution—and emphasizes that these conditions are all indispensable.

This judgment is supported by research data commissioned by Sage from IDC. Financial leaders currently spend more than 12 hours a week validating AI-generated outputs, with 28% of that time consumed by repeated verification and skepticism. Crucially, 70% of CFOs stated they would veto an AI-generated output even if it were 99% accurate, provided they could not see the reasoning behind it. Harris believes this is not CFO resistance to AI, but a clear-eyed recognition of the cost of errors—in the fields of finance and accounting, the cost of missing trust is real and quantifiable.

Cost volatility is another overlooked dimension of trust. Stuart Langer, Senior ERP and Financial & Operational Systems Product Leader at Sage, explains that CFOs are not unable to pay, but cannot stand variability. If the cost is a fixed $2,000 per month, they might dislike it but can accept it; however, $24,000 one month, zero the next, and $24,000 again the following month creates unpredictability, which is the real problem. Reporting author Mark Chillingworth suggests that Sage’s proprietary model, trained on over 15 million invoices, shares similarities with the industry's focus on Small Language Models (SLMs)—the latter being valued by digital decision-makers precisely to reduce the unpredictability of token costs associated with general Large Language Models.

Sage’s response is to make the reasoning process itself a product feature. In AI Workspaces, every conclusion from Sage Copilot includes traceable links showing how the reasoning was derived, with the platform tracking the entire process, involved agents, and task completion methods. Harris summarizes this with three principles: explainable, verifiable, and analyzable. Langer encapsulates the design logic in one sentence: consequences determine control.

The necessary boundary to retain is that all the above arguments come from statements by Sage executives at their own partners’ conference and from research commissioned by Sage. The comparison claiming Sage’s model is more accurate than Anthropic Claude is based on vendor self-testing cited by Langer, which has not been independently reproduced. The IDC data reflects the sentiments of Sage’s customer base and may not represent the entire financial industry. However, as a concrete answer to the question 'why general AI capabilities are sufficient yet specific domains still reject adoption,' this framework of conditions—reasoning visibility, fee predictability, and operational auditability—provides a reference checklist for AI adoption in other industries.

Sources :diginomica

Recherche
Suivant →
Connexion / Inscription
LectureRechercheRadarCadre d'investissement

Griffiths also delineates the applicable boundaries of the skill methodology. Skills are suitable for bounded, repeatable work that benefits from reusable judgment—such as checklists, release processes, migrations, and metric pipelines. If a workflow requires stable interfaces, shared state, audit trails, long-running control, or permissions enforced outside the model, it is better suited to building an application. This distinction reminds readers: skills are carriers for capturing judgment, not a panacea for replacing infrastructure.

Her three practical recommendations boil down to: put business context into a single file for all skills to read; give related skills a consistent path for platform access; use instructions to guide work, and checks and manual approval points to verify work. If a check must block a merge or deployment, that requirement should be configured in repository rules, rather than declared within skill instructions.

Finally, Griffiths offers a pragmatic starting point: begin with the task you are tired of explaining repeatedly. Write down what it needs, where it ends, and what a reviewer sees. Then try it with another agent, improving it as you learn what works.

A continuous evaluation mechanism forms the closed loop enabling this architecture. After each ticket is processed, engineers record actual findings and fixes, and evaluators grade the agent’s response by instrument family as Accurate, Partially Accurate, or Wrong. Results are reviewed to identify underperforming areas; fixes are typically applied by updating the knowledge base, followed by re-running the evaluation set to confirm improvement. In an evaluation of 524 tickets, 269 had corresponding engineer-verified results for comparison. Tickets labeled High Confidence achieved 74% accuracy across 154 cases, while Moderate and Low Confidence tickets achieved 65% across 78 and 37 cases respectively—but the Low Confidence sample size is too small to draw statistical conclusions.

It should be noted that these performance metrics were self-reported by Natera and published by the AWS platform, covering a 90-day production run across 524 tickets. The article does not disclose whether total ticket volume changed during the period, whether selection bias exists (e.g., complex events still relying on human intervention), nor does it provide a controlled experimental design comparing pre- and post-migration states. A 74% accuracy rate for high-confidence diagnoses means that even when the agent is most confident, approximately one-quarter of cases still require independent verification by an engineer.

For teams building agentic systems, two transferable insights stand out from the SWARM case. First, treating skills as versioned production software rather than prompt logic makes the system auditable and locally repairable when errors occur. Second, when accuracy stalls, prioritize checking data coverage over tuning models—the team found that gaps in vendor documentation are the most common cause of missed diagnoses, placing data coverage checks before model optimization. Jirau summarized this priority with his original statement: diagnostic accuracy is primarily a data quality problem.