LesenForschungRadarAnlageframework
Anmelden / Registrieren
AlleTechnikProdukteInvestierenBörseSonstiges

Lesen

2026-10-22168 Beiträge

Aktualisieren

US SaaS Giant Salesforce Announces $5.66 Billion Acquisition of AI Marketing Platform Braze; Shares Surge 51%

Wesentlich
2026-10-21 11:39 GMT+8

US-based enterprise SaaS company Salesforce announced on October 20, 2026, that it intends to acquire AI marketing and customer engagement platform Braze for an enterprise value of approximately $5.66 billion (about RMB 38 billion). Under the agreement, Braze Class A common stock shareholders will receive $47.50 in cash per share. The transaction is subject to shareholder approval and regulatory review, with completion expected in Salesforce's fiscal year 2028.

The acquisition aims to enhance Salesforce's capabilities in real-time customer interaction and personalized marketing. As enterprises increasingly deploy AI agents for marketing and customer service, Salesforce plans to combine Braze's cross-channel marketing and behavioral analysis with its own Customer Relationship Management (CRM) system and Agentforce platform. This integration seeks to help businesses adjust marketing content, channels, and timing based on real-time customer behavior.

Founded in 2011 as Appboy, Braze provides cross-channel customer engagement and marketing automation services. According to its fiscal year 2026 results, the company reported revenue of approximately $738 million, a 24.4% year-over-year increase, though it posted a GAAP operating loss of about $145 million. Braze currently serves around 2,770 customers, including Canva, Instacart, and Venmo.

Following the announcement, market reaction was significant. On October 20, Braze shares closed at $45.57, rising 51.7%, while Salesforce shares remained largely flat at $242.43. The offer price of $47.50 per share represents a roughly 58.1% premium over Braze's closing price on the previous trading day. Salesforce stated that the transaction will not change its fiscal year 2027 guidance and there are no adjustments to long-term targets.

Quellen:zhidx.com

Forschung

GeekPark: From Jev to StartLux, How AI Decision Models Fill the 'Intuition Layer' for Local Deployment

Thema · 决策模型品类Wesentlich
2026-10-04 19:41 GMT+8

In practical AI Agent applications, what occurs densely is often not long-chain deep reasoning, but a multitude of trivial judgments: Which team should handle this ticket? Can this command be released? Is the task complete? Previously, these judgments were either handled "incidentally" by expensive large models or covered by rigid rules—the former slow and costly, the latter unintelligent. TypeSafe AI's Jev model proposes a new architectural approach: stripping "judgment" from "generation" to create a cheap, high-frequency infrastructure. This type of model, referred to as "System One," only answers questions with preset options, directly returning structured results with probabilities at a very low price point.

Jev's success proved the market value of this path, but its closed-source hosted nature limited its application in scenarios requiring data sensitivity or local deployment. This left an entry point for followers. Recently, global giants including OpenAI, Cloudflare, and Amazon have followed suit with similar APIs or models. Meanwhile, Chinese teams rapidly adopted a differentiated strategy: open-sourcing and localization. Shanghai AI Lab open-sourced the multimodal decision model Intern-Decision, while StartLux (Yuan Dian Xing Hui), a Shanghai company founded just five months ago, released StartLux-Decision with support for quantization files, offering specifications ranging from 0.8B to 27B, aiming to keep model capabilities and data on users' own devices.

For companies betting on "Local AI," decision models are not a hot new category to chase, but a critical component for whether local systems can run effectively. Cloud computing power can be stacked, but personal computers have hard limits on VRAM, memory, and power consumption. If every small judgment requires a 27B parameter large model, local Agents simply cannot run efficiently. Data published by StartLux shows that its 4B version, after Q4 quantization, requires only about 2.71GB of VRAM, maintaining a 98.3% consistency rate with original weights. This layered architecture—large models handling general capabilities, small models handling high-frequency judgments—makes building a complete personal AI system on resource-constrained devices possible.

However, the moat in this field currently appears shallow. Jev claimed two years of research, yet OpenAI produced a similar product in two weeks, StartLux in three days, and Cloudflare caught up within half a month. A category that can be replicated in two weeks is difficult to form a long-term technical barrier. StartLux CTO Guo Quanwei also acknowledged that the real challenge lies not in the model itself, but in long-task stability, exception recovery, and overall experience. When "judgment" becomes cheap and ubiquitous, value may ultimately fall not on the models themselves, but on those who can assemble these low-cost components into reliable products.

Quellen:GeekPark

Forschung

US Confidential AI Startup Opaque Systems Announces Team Joins Anthropic and Ceases Operations

Kurzfassung
2026-10-22 08:00 GMT+8

According to a report by iThome, US-based confidential AI and verifiable AI technology startup Opaque Systems announced on October 13 that its team will join AI company Anthropic, while the company itself will cease operations.

The report states that Anthropic has acquired a technology license from Opaque, and some employees will transfer to Anthropic, including Opaque CEO Aaron Fulkerson and CTO Rishabh Poddar. The transaction amount and full terms have not been disclosed.

Opaque originated from the RISELab research team at UC Berkeley in the United States, focusing on using Confidential Computing technology to protect sensitive data during AI model execution. As AI applications developed, its research direction expanded to Verifiable AI.

Opaque stated that related open-source work will continue through AgenTrust, including projects such as Confidential A2A and TRACE.

Quellen:iThome

Forschung

US NIST, NSA and CISA Release Voluntary Guidelines for AI Agent Incident Response

Wesentlich
2026-10-22 08:00 GMT+8

On October 16, the U.S. National Institute of Standards and Technology (NIST), the National Security Agency (NSA), and the Cybersecurity and Infrastructure Security Agency (CISA) jointly released 10 recommendations for AI agent incident response. This voluntary guidance primarily targets organizations training and evaluating highly advanced AI agents, though it also serves as a reference for those deploying such agents. According to iThome, NIST divided the content into five stages: preparation, detection, response, recovery, and notification.

The non-profit cybersecurity organization Cloud Security Alliance (CSA), in its interpretation of the guidelines, noted that organizations must not only monitor the tools used by AI agents, their system operations, and network activities but also guard against agents interfering with monitoring tools or tampering with logs to ensure effective anomaly detection. To reduce the risk of boundary violations, the guidelines suggest that organizations pre-limit agent access permissions and establish emergency stop mechanisms; if an agent violates security limits, its activity can be immediately halted and permissions revoked. Additionally, organizations should maintain complete records of AI agent tool calls, including blocked operations, to facilitate post-incident reconstruction.

After an incident, organizations need to re-test against previously identified boundary violations to confirm that adjusted protections prevent the same operations before resuming agent functionality. NIST also recommends that organizations fulfill legal notification obligations, providing event information to government agencies or publicly disclosing incidents as necessary.

Quellen:iThome

Forschung

Perplexity Open-Sources SPACE: AI Agent Sandbox Passes 108 Escape Tests in Self-Assessment

Thema · 智能体沙箱隔离Wesentlich
2026-10-22 08:00 GMT+8

AI search company Perplexity announced on October 22, 2026, that it is open-sourcing its AI agent sandbox platform, SPACE (Sandboxed Platform for Agentic Code Execution). The platform aims to provide an isolated code execution environment for AI agents, reducing the impact of running untrusted code on host systems. Previously, SPACE was used for Perplexity's own AI agent service, Computer, replacing commercial sandbox services in July of this year.

Technically, SPACE employs Firecracker microVM technology to ensure each AI agent runs in an independent virtual machine with its own operating system and writable file system, fully isolated from the host. A control layer outside the sandbox manages VM creation, network access, and platform credentials. This design allows AI agents to access authorized services while preventing sensitive credentials from being exposed inside the sandbox.

The Perplexity research team disclosed security test results for SPACE: 108 virtual machine escape tests were conducted with no successful escapes found. However, during testing, they discovered that AI models could bypass certain network access restrictions. The team stated these issues have been fixed, and subsequent tests confirmed models failed to bypass them again. The source code is now public on GitHub under the Apache 2.0 license. Akamai has tested and launched SPACE on its cloud platform, CoreWeave provides deployment guides, and Crusoe is collaborating to integrate it into its AI service portfolio.

Quellen:iThome

Forschung

Anthropic Launches Claude Duet: A Proactive AI Agent Preview for US Pro/Max Users

Strukturell
2026-10-21

Anthropic launched a research preview of its new application, Claude Duet, on October 21, 2026. Unlike traditional passive chatbots, Claude Duet is designed as a "proactive agent" capable of initiating contact with users based on data from connected calendars, emails, and apps. It sends daily briefings and continues working on tasks even when the user steps away. Currently, access is limited to US Pro and Max subscribers who join a waitlist.

The core functionality of Claude Duet lies in its judgment. Built on the Claude Opus 5.5 model, it aims to help users prioritize what needs immediate attention versus what can wait. For instance, it sends a morning briefing across connected apps and proactively notifies users of urgent events like schedule conflicts or unpaid invoices. Users can delegate complex tasks ranging from scheduling meetings to building dashboards and creating financial forecasts.

To address the security and privacy risks inherent in proactive agents, Anthropic implemented multiple layers of protection. First, Claude Duet uses a dedicated classifier and safety system prompt to prevent the leakage of sensitive information such as credit card numbers and government IDs. Second, all critical actions—including sending messages, making payments, or signing documents—require explicit user approval before execution. Additionally, the system integrates defenses against prompt injection attacks and has undergone red teaming and external penetration testing.

Technically, Claude Duet operates within its own virtual machine and browser environment, allowing it to work continuously without relying on the user's device being online. It checks connected account statuses to trigger actions, though these checks occur every few hours rather than continuously to conserve resources. This design balances responsiveness with computational cost.

Although still in preview, Anthropic has demonstrated potential commercial applications. Internal employees and early testers, including small business owners, reported that Claude Duet effectively handles routine operational tasks such as accounting, customer support, and marketing. However, given the limited nature of this preview, long-term stability and broader user acceptance remain to be seen.

Quellen:Claude

Forschung

Morgan Stanley Podcast: Asian Bonds Need a Bigger Map

Wesentlich
2026-10-22 03:57 GMT+8

Morgan Stanley has released a new episode of its financial podcast 'Thoughts on the Market' titled 'Asian Bonds Need a Bigger Map'. Recorded in Hong Kong on October 13, 2026, and published on October 21, the episode features Kelvin Pan, Head of Asia Credit Strategy at Morgan Stanley.

Pan argues that while Asia-Pacific fundamentals remain strong, the traditional investment universe for Asian bonds faces increasing constraints. Since early 2023, US policy rates exceeding those in many Asian jurisdictions have reversed the environment supporting Asian USD credit, leading to shrinking supply and elevated valuations. Concurrently, assets in funds dedicated to Asian credit and high-yield debt have significantly declined from their 2021 peak.

To address this, Pan suggests the issue lies not with Asia itself but with how investors view it. He introduces the concept of 'APAC Bonds', which expands beyond traditional USD credit and emerging market local currency government bonds to include convertible bonds, local currency corporate bonds, and coverage of Japan and Australia. This broader map aims to help investors flexibly manage risk and return across different economic environments by combining diverse asset classes.

Quellen:youtube.com

Forschung

OpenAI Releases V7 Go Video: Showcasing AI Agents Optimizing Investment Screening and Insurance Underwriting via Context Graphs

Thema · OpenAI V7 GoWesentlich
2026-10-22 06:05 GMT+8

OpenAI has published a video introduction to its V7 Go product on its official YouTube channel. According to the video description, the core functionality of V7 Go is organizing unstructured company files into a "Context Graph," enabling AI agents to directly query this structured information to perform specific tasks.

The video specifically emphasizes applications in vertical financial fields, including screening investment deals and supporting insurance underwriting processes. The description cites specific performance data: on V7's hardest graph-query tests involving thousands of documents, a model named "GPT-6 Astra" reached 89% accuracy. Additionally, the video title claims the solution reduces costs by 78%.

For readers interested in the actual efficacy of AI in professional workflows, this resource provides a concrete case reference, demonstrating an attempt to transform general large models into domain-specific knowledge graph constructions. Note that all mentioned data originates from OpenAI's video description, representing vendor-reported results that have not yet been independently reproduced or verified through detailed technical white papers.

Quellen:youtube.com

Forschung

Microsoft Open-Sources Scope to Evaluate and Trace AI Coding Agents

Thema · 微软Scope评测平台Wesentlich
2026-10-20 00:00 GMT+8

On October 19, 2026, Microsoft open-sourced Scope, an evaluation platform designed to help development teams test and track the performance of AI coding agents such as GitHub Copilot and Claude Code when executing specified tasks. The platform allows software teams to verify whether agents correctly use Software Development Kits (SDKs), Application Programming Interfaces (APIs), and related tools according to documentation, while recording the entire operation process and test results.

In practice, AI agents assisting developers can make mistakes such as selecting the wrong API, ignoring important instructions in documentation, or failing to complete work after encountering errors. Even if the final generated code runs successfully, it does not guarantee that the agent used the expected toolchain. Scope addresses this pain point by enabling teams to test error-prone aspects of agent interactions with proprietary SDKs, command-line tools, and Model Context Protocol (MCP) servers used to connect external services.

Development teams can define completion criteria for tasks and provide starting code, allowing different AI agents to execute the same task. The system records the AI model version, agent version, and tool configuration for each run. After modifying documentation or tools, teams can re-test with the same task to observe if agents complete work more easily and distinguish between sporadic and recurring errors. Additionally, Scope supports setting multiple check conditions with dependencies (e.g., successful build before checking SDK initialization) and uses large language models to assess condition fulfillment, preserving operation logs and directory snapshots for troubleshooting.

Despite automated assessment, Microsoft notes that AI agents may produce different results even with identical settings, and automatic judgments can be erroneous. Therefore, the company recommends repeated testing and human review of records. Currently, Scope requires user deployment, offering a web management interface and API integration options. The project is MIT-licensed, source code is available on GitHub, and it remains in a research preview stage.

Quellen:Microsoft Open Source Blog

Forschung

German Software Giant SAP Reports Record Q3 Revenue, CEO Attributes Cloud Growth to AI Agent Joule

Wesentlich
2026-10-22 09:01 GMT+8

SAP SE, Europe’s largest software company and a German enterprise application giant, reported record-breaking quarterly revenue on October 21, 2026, surpassing the €10 billion barrier for the first time with total sales of €10.27 billion ($11.2 billion). Adjusted earnings per share came in at €1.89, beating analyst expectations of €1.83.

Cloud operations served as the primary growth engine: SAP’s cloud revenue rose 25% year-over-year to €6.67 billion, while current cloud backlog surged 26% to €24.11 billion, exceeding the 24%-25% growth forecast by analysts. Non-IFRS operating profit climbed 18% to €3.03 billion, and free cash flow increased 36% to €1.72 billion.

Chief Executive Christian Klein stated clearly during the conference call: “AI was the main driver behind customer’s move to the cloud during the quarter.” He revealed that hundreds of customers have signed up for new subscriptions to access SAP’s autonomous AI agent tool, Joule. Designed as a conversational interface layer connecting to more than 70 AI models, Joule acts as a core component of the “autonomous enterprise” strategy, working alongside existing SAP applications rather than replacing them.

Chief Financial Officer Dominik Asam noted that part of the cloud growth stemmed from customers recovering from previous financial distress. While raising the full-year subscription revenue guidance toward the upper end of the €25.8 billion to €26.2 billion range, Asam warned of a gloomy economic backdrop, citing rising oil prices and U.S. interest rates that are “starting to develop teeth.” The fourth quarter will be critical in determining how the rest of the fiscal year lands.

Quellen:SiliconANGLE

Forschung

US Army Solicits ECHOTRIBBLE Prototype: AI Agents to Autonomously Navigate and Neutralize Digital Threats

Thema · 美陆军AI网络代理Wesentlich
2026-10-22 08:38 GMT+8

According to Tom's Hardware reporting and Govly's summary of public announcements, the US Army is developing a program named ECHOTRIBBLE, publicly soliciting prototypes for modular autonomous AI agents. The initiative is led by the US Army Cyber Fusion Innovation Center (CFIC) in collaboration with Army Cyber Command (ARCYBER) and the Army Cyber Technology and Innovation Center (ArCTIC), inviting industry submissions of white papers.

Solicitation W519TC-26-AE-ECHO was posted on July 16, 2026, with a response deadline of August 14. Official language uses "neutralize" rather than the more aggressive "seek and destroy" found in some media headlines. The stated design purpose is to "autonomously navigate infrastructures, identifying and neutralizing identified threats/targets." Key requirements include modularity, low resource consumption, secure communications, and multi-platform deployment, focusing on software solutions without specifying particular hardware or vendors.

The project remains at the prototype solicitation stage, with no disclosed winning vendor, deployment timeline, or detailed rules of engagement. Per Govly's summary, online demonstrations took place from October 6 to 8, with final selection and award expected in mid-October. If successful and funded, subsequent production contracts may follow. Experts warn that the real-world consequences of such autonomous AI agents in digital environments could be severe, and the specific scope of "neutralization" awaits further official clarification.

Quellen:INSIDE

Forschung

US Judge Orders Meta to Pay $25 Billion Penalty in New Mexico Cambridge Analytica Case

Thema · Meta剑桥分析罚款案Strukturell
2026-10-22 01:53 GMT+8

US District Judge Francis Mathew in New Mexico ordered social media giant Meta to pay a $25 billion civil penalty on October 21, 2026, for deceiving users about data privacy in connection with the Cambridge Analytica scandal. This ruling marks a new scale for state-level enforcement actions against major technology companies regarding historical data violations.

The case stems from revelations in 2018 that Cambridge Analytica, a British political consulting firm, improperly harvested personal data from up to 87 million Facebook users via a third-party app to profile voters during the 2016 US presidential election. New Mexico Attorney General Raul Torrez sued Meta in 2021, alleging violations of state consumer protection laws. A jury found in September that Meta committed more than 43 million violations, each carrying a statutory penalty of up to $5,000.

Although New Mexico law allowed for penalties as high as $219 billion, and state lawyers sought between $35 billion and $40 billion, Judge Mathew set the final amount at $25 billion. Vincent Joralemon, a Senior Fellow at the Berkeley Center for Law & Technology, noted that while the finding of liability is likely to stand on appeal due to damning evidence, the actual damages might be reduced.

Meta Chief Legal Officer CJ Mahoney stated that the company will appeal, arguing the lawsuit seeks to impose a "tax on speech." This decision adds to a series of costly legal setbacks for Meta, which recently agreed to an $18 billion settlement with 48 states over teen social media addiction and faces additional penalties in another New Mexico case involving child safety.

Quellen:New York Post

Forschung
Weiter →
Anmelden / Registrieren
LesenForschungRadarAnlageframework