LesenForschungRadarAnlageframework
Anmelden / Registrieren
AlleTechnikProdukteInvestierenBörseSonstiges

Lesen

2026-10-27149 Beiträge

Aktualisieren

CFA Institute Podcast: Cindy Lam Shares Journey from Equity to ESG Research

Kurzfassung
2026-10-27 20:21 GMT+8

CFA Institute has published the second episode of its podcast series "What Do They Actually Do?", focusing on the role of an ESG analyst. The guest is Cindy Lam, a Chartered Financial Analyst (CFA) and ESG investment professional based in Hong Kong. In this episode, she discusses her career path, specifically detailing her transition from traditional equity research into the specialized field of ESG research.

The publisher's description highlights several core competencies that investment professionals need to develop for a successful career in ESG. These include the ability to analyze both financial and sustainability data, monitor regulatory and market developments, and develop thematic research with investment insights. Additionally, the episode emphasizes the importance of presenting ideas to clients, applying valuation and modeling skills, and building strong communication abilities.

For professionals considering a move into ESG or sustainable investing, this resource offers a concrete overview of the day-to-day responsibilities and skill sets involved. Cindy Lam also shares personal lessons from her career, such as the value of asking questions, learning from experienced colleagues, and thinking independently to form and defend investment views. The episode aims to clarify how sustainability considerations are influencing investment research across Asia.

Quellen:youtube.com

Forschung

UK Robotics Startup Machenta Raises $40M Seed Round, Claims Simulation-Only Robot Training

Thema · Machenta融资Kurzfassung
2026-10-27 18:00 GMT+8

London-based robotics startup Machenta has raised a $40 million seed round co-led by Creandum and Union Square Ventures, with valuation undisclosed. Founded by researchers from Imperial College London, the company aims to build an AI model that teaches robots new physical tasks after watching them performed just once. CEO Edward Johns said the model trains entirely in computer-generated environments without using real-world data. The startup has 11 employees and no near-term commercialization plans.

Quellen:Business Insider

Forschung

Backup Vendor Veeam Launches Tool to Track and Roll Back AI Agent Changes in Microsoft 365

Thema · VeeamWesentlich
2026-10-27 18:47 GMT+8

US-based data protection and backup software company Veeam Software has announced the general availability of its Intelligent ResOps suite for Microsoft 365. The tool addresses growing enterprise concerns about how to precisely identify and restore specific changes when autonomous AI agents accidentally modify or delete data within SharePoint, OneDrive, Teams, and Exchange.

Veeam CEO Anand Eswaran cited a recent case handled by the Spanish Data Protection Agency (AEPD) as context for the feature's development, where a third-party deployed AI agent accessed and modified personal data without authorization. Under regulations like the EU General Data Protection Regulation (GDPR), companies must report data breaches within 72 hours, requiring them to know not just that changes occurred, but exactly which records were altered, their current state, and their original state. Traditional whole-system recovery points cannot provide this granularity. Veeam's new tool uses a "recovery map" concept, allowing administrators to view a timeline of agent actions and selectively roll back individual files or datasets.

The product is built on Veeam's DataAI Command Graph platform, linking production data with backup context to map files, users, permissions, agents, and protection status. According to the vendor, it detects known and unknown AI agents (including Microsoft Copilot) and shows which sensitive data they accessed and which files they modified or deleted. It can also quarantine redundant, obsolete, or trivial (ROT) data. Pricing is set at $2.49 per user/month for the Foundation edition and $4.99 for the Advanced edition.

Beyond Microsoft 365, Veeam announced private preview plans for identity provider Okta and code repositories (GitHub/Azure DevOps). Okta lacks native point-in-time recovery, with only a 30-day soft-delete window for users. Veeam claims its service creates immutable recovery points for 29 object types and correctly restores memberships, app assignments, and roles during resets. These extended features are currently in testing, and final capabilities may change.

Quellen:iX Magazin

Forschung

Zvi Mowshowitz: Anthropic's Safety Sacrifices Favor Founder Concerns, Failing to Address Loss-of-Control Risks

Thema · AnthropicWesentlich
2026-10-27 20:49 GMT+8

In a blog post dated October 27, 2026, Zvi Mowshowitz presented a core observation regarding Anthropic: the company simultaneously does things that are admirable and things that are disappointing, and these are not randomly distributed—they split along the lines of CEO Dario Amodei’s personal concerns. Mowshowitz opened with the statement: 'Anthropic is a frustrating company.'

When responding to investigative journalist Nathan Calvin’s criticism of Anthropic’s leaked draft bill, Dario Amodei listed nine safety actions taken over the past year that carried substantial commercial costs. These included advocating for SB 53 legislation, implementing strict biological safety filters, delaying the release of Mythos-class models, publicly arguing that governments have the right to block model releases, supporting export controls on chips to China, restricting AI service sales to companies controlled by the People’s Republic of China (PRC), publishing numerous risk reports, and dedicating 5% of shares to a public benefit fund.

However, Mowshowitz noted a common thread among these sacrifices: they are concentrated in areas where Dario personally cares deeply—specifically biosecurity, export controls to China, and restrictions on Chinese companies. He cited tech journalist Kelsey Piper’s observations from an internal Anthropic briefing as counter-evidence. When asked about plans for superintelligence alignment and loss-of-control risks, Piper summarized the response as: 'No plan, no plan, no plan'—meaning there was no plan for superintelligence alignment, no plan to avoid concentration of power or loss of control, and no vision for what a good world with superintelligence would look like.

Cody Fenwick further pointed out that while Dario claims Anthropic’s strategy has 'always been' to pursue the highest standards, in a 2022 interview Dario explicitly stated that Anthropic 'shouldn't be racing ahead,' suggesting models only needed to be 'reasonably close to state of the art.' This positioning differs significantly from 'pursuing the highest standards.'

Mowshowitz argues that this selective pattern has concrete consequences at the governance level. Anthropic has filed its S-1 document with the U.S. Securities and Exchange Commission (SEC). The company recently announced the establishment of a Public Benefit Fund, committing 5% of its shares via new Class N stock (which carries economic interest but no voting rights)—subject to shareholder approval—and pledging an additional 2% at the board’s discretion for every 50% increase in company value above a set baseline.

Independent evaluation agency METR has obtained 'employee-like access' to Anthropic, allowing it to retrieve most categories of internal information, with exceptions only for customer data, attorney-client privilege, and key intellectual property such as model architecture details. METR can publish its analysis of Anthropic’s risk levels, incidents, or practices. While Anthropic cannot veto publication, it can require redaction of information falling into specific, predefined categories.

Quellen:thezvi.substack.com

Forschung

Unity reports 98% of its CLI executions are driven by AI agents, with Claude Code leading at 60%

Thema · Unity CLIWesentlich
2026-10-27 16:35 GMT+8

Game engine company Unity published cumulative usage data for its command-line interface tool, "Unity CLI," on October 27. The figures reveal that between July 20 and October 13, 2026, 98% of all executed commands originated from AI coding agents and other automation processes rather than direct human input. Only 0.2% of commands were typed manually into the terminal, while 0.3% were run through CI pipelines.

Unity CLI is a single-binary tool that allows users to install and manage the Unity engine within a terminal environment. When integrated with pipeline packages, it can control active Unity editors, enabling scripts, continuous integration systems, and AI coding agents to access the development environment in the same way. This structure suggests that after developers complete initial setup, the bulk of subsequent detailed work is led by AI agents.

Among the AI coding agents linked with Unity CLI, Anthropic's Claude Code accounted for 60% of usage, followed by OpenAI's Codex at 26%, with Cursor and Grok also present. The majority of this volume concentrated following the release of official plugins for these tools. Unity stated it will continue to support agents without official plugins to perform editor operations via the CLI.

The statistics are based on Unity CLI telemetry, excluding Unity employees and users who opted out of data collection. Over 226,000 logged-in users generated more than 200 million execution events. By country, the United States, Japan, and South Korea ranked as the top three markets.

Quellen:zdnet.co.kr

Forschung

User Reports Claude Opus 5.5 Blocked Request to Reverse Engineer LG Smart TV for Ad Removal

Kurzfassung
2026-10-27 21:54 GMT+8

According to heise online, a Reddit user reported that Anthropic's Claude Code tool refused to assist in reverse engineering an LG Smart TV. The user aimed to use the AI to analyze the television's software to remove pre-installed apps, block advertisements, and limit data tracking. Screenshots indicated that the automatic safety mechanisms of the Claude Opus 5.5 version classified the request as potentially problematic and blocked it.

The incident was first reported by TechCrunch and subsequently covered by heise online. It remains unclear whether this represents a systematic prohibition on assisting with consumer hardware reverse engineering or an isolated false positive. Anthropic has not issued an official statement regarding this specific case.

Reverse engineering is often used in the smart home sector to enhance device privacy and control, though it carries risks such as bricking devices through improper modifications. This case highlights the difficulty AI assistants face in distinguishing between legitimate 'modifying owned devices' intents and potential malicious tampering, serving as a notable usability signal for users relying on AI for personal tech customization.

Quellen:heise online

Forschung

EU Agency ENISA Releases Software Bill of Materials (SBOM) Technical Guide: Proposes Six-Stage Lifecycle and Anti-Tampering Strategies

Thema · ENISA SBOM指南Wesentlich
2026-10-27 15:19 GMT+8

The European Union Agency for Cybersecurity (ENISA) officially released the "Technical Advisory for Software Bill of Materials" on October 22, 2026. The guide aims to enhance software supply chain transparency and assist companies in meeting obligations under the EU Cyber Resilience Act (CRA). It proposes a six-stage lifecycle management process covering generation, validation, signing, storage, usage, and maintenance, recommending integration into Continuous Integration/Continuous Delivery (CI/CD) pipelines.

ENISA specifically highlights security risks associated with SBOMs, using the STRIDE threat model to analyze six categories of threats: spoofing, tampering, repudiation, information disclosure, denial of service, and elevation of privilege. Attackers could impersonate legitimate vendors to distribute tampered SBOMs, causing enterprises to mistakenly trust malicious or compromised components. Without proper access controls, version and dependency information within SBOMs could also leak, providing intelligence to attackers. The guide recommends digital signatures, identity verification, access control, and audit logs to strengthen SBOM integrity.

Addressing practical implementation challenges, ENISA cites an industry survey from June showing that while 41% of respondents want fully automated SBOM generation during each build, only 26% expect full automation during the product support period. This underscores the need for continuous update, validation, and version tracking mechanisms to prevent discrepancies between the list and actual software versions. Additionally, the guide lists available open-source tools categorized by CycloneDX and SPDX formats, such as Syft, cdxgen, Cosign, and Trivy, offering references for micro and small-to-medium enterprises.

ENISA explicitly states that this guide is a non-binding technical recommendation that does not create or change legal obligations and should not be viewed as an official legal interpretation of the CRA. However, its practical advice holds significant reference value for software vendors seeking to proactively prepare for compliance.

Quellen:iThome

Forschung

OpenAI Announces Three AI Scientific Breakthroughs and Introduces 'Shepherd' Verification Mechanism

Thema · OpenAIStrukturell
2026-10-26 08:00 GMT+8

On October 26, 2026, OpenAI announced that its internal models achieved three significant discoveries in mathematics, computer science, and physics: a proof related to the local smoothing conjecture in harmonic analysis (implicating the Kakeya conjecture), a refutation of the Exponential Time Hypothesis (ETH) in computational complexity, and an existence proof for four-dimensional Quantum Chromodynamics (QCD). OpenAI characterized these outputs as "raw research data" and introduced a novel release strategy inviting external experts to act as "shepherds," organizing groups to understand and explain the results.

For the computer science result concerning the ETH refutation, OpenAI formed a collaboration group led by Russell Impagliazzo, Ramamohan Paturi, and Ryan Williams. Impagliazzo and Paturi are the original proposers of the ETH, while Williams is a leading expert in the field. OpenAI committed to making all raw data public to facilitate independent verification and pledged to fund a conference to explore the broader implications for algorithm design and complexity theory. For the mathematical result on local smoothing, OpenAI chose to directly release the AI output and Lean formalization proofs, leaving researchers free to organize their own efforts to explain them. In the physics domain, OpenAI deferred the release timing to the external collaboration group.

This initiative aims to address the trust deficit surrounding AI-generated scientific claims. By adhering to principles of transparency, no academic credit for OpenAI, recognition for contributors, and independence, OpenAI seeks to establish a new paradigm where human experts translate AI's raw outputs into accepted scientific knowledge. Although these results have not yet undergone full peer review and their correctness remains subject to community validation, this "human-AI collaborative verification" model may serve as a crucial reference for future AI involvement in frontier scientific research.

Quellen:OpenAI

Forschung

Voice AI firm Regal launches Goal Oriented Agents with actor/director model split

Thema · Regal VoiceKurzfassung
2026-10-27 22:00 GMT+8

US-based enterprise voice AI company Regal Voice Inc. announced on October 27, 2026, the addition of "Goal Oriented Agents" to its platform. This feature allows businesses to define specific objectives for calls and employs two AI models working together to achieve them.

The design targets a common issue in traditional conditional-logic voice agents, such as those used for auto insurance claims: when a caller provides multiple details at once, older systems often circle back to ask for information already given just to fill each field. Regal states that building an agent on a single prompt keeps conversations open but leaves business processes loosely followed.

The new agents deploy two models per call: an "Actor," a smaller, faster model tuned for dialogue that replies immediately, and a "Director" running behind it to check if current stage goals are met and move the call to a new stage with fresh instructions. Businesses build agents from "goal nodes," listing required outcomes and prerequisites (like date of birth or mentioning a death in the family), which can be written in plain language.

According to SiliconANGLE, Regal identifies complex intake and qualification calls as strong use cases, including Medicare enrollment, roadside assistance, and mortgage qualification. The company serves over 200 enterprises, including Google LLC, AAA, and Toyota Motor Corp. Regal is a venture-backed startup that has raised $83 million to date, including a $40 million round in October 2024.

Quellen:SiliconANGLE

Forschung

Tim Tyler of Gravitee Proposes Agent Accountability Framework: Governing AI Agents as Digital Workforce

Thema · AI代理问责框架Wesentlich
2026-10-27 17:00 GMT+8

As enterprises scale their deployment of AI agents, a core question emerges: who is responsible when an agent acts autonomously? Tim Tyler of Gravitee addressed this at the AIE CODE summit by proposing the Agent Accountability framework. He argues that AI agents should not be treated merely as software tools but managed as a 'digital workforce,' structured around three critical stages.

The first stage, 'Assignment,' requires establishing a clear identity, team affiliation, and a designated 'human sponsor' for each AI agent. This resolves the accountability vacuum created by anonymous agents. The second stage, 'Alignment,' focuses on defining the agent's role and mission while strictly enforcing the 'principle of least privilege,' ensuring agents possess only the minimum permissions necessary for specific tasks to prevent unauthorized actions. The third stage, 'Administration,' introduces oversight scorecards and 'guardian agents' to provide step-by-step recourse when agents deviate from their intended path.

Tyler notes that many companies currently deploy AI agents in a manner akin to hiring employees with no ID, no manager, and shared passwords—a model fraught with risk. His framework aims to bridge the gap between technical capability and managerial accountability through structured governance. For organizations exploring the scaled application of AI agents, this perspective offers a systematic reference ranging from identity authentication to permission control and continuous supervision.

Quellen:youtube.com

Forschung

Oracle Fusion Claw Architecture: Consumes AI Units Only During Reasoning, Delegates Execution to Deterministic Computing

Thema · Fusion ClawWesentlich
2026-10-27 18:03 GMT+8

When Oracle's Fusion Claw was released in late September 2026, the public already knew it used large language models for reasoning and traditional computing for execution. However, the specifics of this separation and the resulting cost reductions were explained more fully by Chris Leone, Executive Vice President of Applications Development at Oracle, during his session at the AI World conference on October 27, 2026.

Leone stated that the economic shift brought by Fusion Claw comes from two levels: within individual results and across results. Within a single result, the core principle is to invoke AI units only when reasoning is required—using large models to create plans, assemble tools, and generate execution logic—and then hand off subsequent work to deterministic computing. The reasoning phase may consume AI units (depending on the model used), but the 'deterministic computing' and 'precision execution' phases do not require large models and do not consume AI units.

Leone cited a customer example: a global short-term rental platform that writes approximately one billion records to its general ledger each month, needing to understand the start and end times of each stay to correctly recognize revenue. To solve such high-volume scenarios with AI, the key challenge is scaling economically and precisely. His answer was to complete the processing of millions of records, thousands of calculations, and hundreds of scenario simulations outside the large model, ensuring consistent, precise, and cost-controllable results every time.

The mechanism at the cross-result level is particularly noteworthy. Leone said Oracle's engineering team takes 'notes' after each result is completed—reviewing the result, plan, and context to record which path was taken. When encountering a similar path next time, the system does not need to rediscover all branch scenarios. He claimed internal tests show this saves up to 50% of subsequent reasoning time. As noted by Jon Reed in his reporting, this 'learning' occurs almost entirely outside the large model, because learning within a large model implies costly retraining.

Regarding customer-side cost control, Leone introduced several specific tools. AI Studio includes a budget and control plane where budgets can be set at the macro-level, individual agent level, and agentic application level, with real-time termination triggered upon reaching thresholds. A 'kill switch' allows shutting down specific agents. The CLI testing framework allows customers to run agents for a period before formal deployment to simulate costs and generate estimated budgets.

Most notably, there are model downgrade recommendations. Leone said Oracle supports downgrading from the latest Luna or OpenAI models all the way to version 4.1, and also supports open-weight models like gpt-oss—which is free for customers and runs on OCI. The optimization loop in the testing framework informs customers which agents or workflow steps can be downgraded to cheaper models without losing accuracy. He stated this capability has existed since 2016 and customers have been using it.

Quellen:diginomica

Forschung

Insurance Tech COO: The Hardest Problem with AI Agents Isn't Technology, It's Accountability

Thema · AI智能体问责制Wesentlich
2026-10-27 19:00 GMT+8

A landlord in California has 21 days after a resident moves out to return the security deposit or send an itemized list of deductions. Miss it in bad faith, and the penalty can reach twice the deposit. That deadline doesn’t care whether the letter was supposed to come from a person or a piece of software. Either way, someone will answer for it.

Rick Folgmann is the chief operating officer at Get Covered, an insurance technology company. His company now runs 16 AI agents across more than 3 million rental units. Recalling the experience of building these agents, he said that technical questions turned out to be the easier part; the truly difficult question for every agent is who it reports to.

He believes most companies are about to face the same issue, yet few are prepared. Deloitte’s 2026 State of AI survey found that 23% of enterprises currently use agentic AI at least moderately, and 74% expect to reach this level within two years. Meanwhile, Gartner predicts that more than 40% of agentic AI projects will be canceled by the end of 2027.

Folgmann’s conclusion is that these figures indicate agents should operate like employees in regulated roles—with written rules, defined scope, a manager, and regular reviews. He offers three practices distilled from his own operations.

First, start with tasks that already have rules. The best early candidates for agents are tasks where rules already exist outside anyone’s head: a statute, a contract clause, or insurance requirements in a vendor agreement. He cited the example of verifying vendor certificates of insurance—property teams across the industry spend roughly three hours per property every week chasing these documents via email. The work is tedious but completely specified. Agents have objective standards to validate output, and when the team audits results, there is no need to debate judgment calls. His rule of thumb is: if a new employee cannot make the decision according to your description on their first day, neither can an agent.

Second, set autonomy by workflow. He noted that tech executives often ask whether an AI system is autonomous, as if it were a fixed attribute of the product. In reality, the same agent should operate at different levels of freedom depending on what it touches. At Get Covered, a reminder to a vendor whose certificate expires next month can be sent automatically; a response to a resident requesting an emotional support animal never will. He advises leadership to list every workflow an agent will touch and ask the same question for each: If this action is wrong, how long until we discover it, and what will the cost be by then? Quick detection and low cost earn the agent more room.

Third, assign a named human owner to each agent. Folgmann learned early in his career in the financial industry that every automated credit decision must eventually be explained to a customer or regulator.

The model’s output has never been an acceptable answer. This principle still holds, except agents now act more frequently and across more systems than any underwriting model did. Therefore, at Get Covered, every agent has a person ultimately responsible, just as a team or a budget has an owner. That person reviews its error logs, decides when to expand or restrict its autonomy, and takes responsibility for it.

Quellen:Crunchbase News

Forschung
Weiter →
Anmelden / Registrieren
LesenForschungRadarAnlageframework

Mowshowitz also acknowledges clear boundaries to his own analysis. He wrote: 'I know more about Anthropic than almost anyone, yet I cannot phrase my worries about it in a way that rings true to people who work inside—it may be that I am wrong, or perhaps because they are too busy and too tired.' He further noted that it remains unclear whether Anthropic’s leadership agrees that there is a genuine problem here.

For readers evaluating Anthropic as an investment in the AI industry, Mowshowitz’s framework provides a specific test standard: do not just look at what the company claims to care about, but examine which areas it is willing to bear quantifiable commercial costs in, and whether those costs cover the most severe threats within the company’s own risk narrative.

Fusion Claw currently has three operating modes: full auto, research, and assist. Full auto operates autonomously within a governance framework, while assist acts as a co-pilot advisory layer. AI unit consumption may be influenced by the mode, but use case complexity and the number of loops will deterministically affect consumption.

It must be clarified that all the above cost data are internally tested results reported by Oracle. As pointed out in SiliconANGLE's report in late September, Oracle did not provide measured cost savings data or production environment performance evidence. The 50% reduction in reasoning time, accuracy retention after model downgrades, and the actual effectiveness of the 'note-taking' learning mechanism remain at the level of vendor claims.

However, from an industry investment perspective, this architectural explanation itself holds independent value. It indicates that Oracle believes the scaling bottleneck for enterprise-grade agentic AI is not model capability, but reasoning cost. Limiting large models to 'planning and assembly' roles, handing repetitive calculations to deterministic engines, and reducing repeated reasoning through cross-task learning—if this approach holds true, it means the competitive dimension of enterprise AI platforms is shifting from model performance to cost architecture efficiency.

Jon Reed, a reporter for diginomica, pressed Leone on real-time cost monitoring after the session. Leone admitted that real-time cost monitoring is still under development, but existing budget controls, kill switches, and simulation frameworks already provide real-time constraints. This suggests Oracle itself realizes that for 'always-on' agent loops, customers care less about upper limits of capability and more about whether the lower limits of cost are controllable.

He also points to a quieter problem: when multiple teams launch agents independently, no one sees where those agents contradict one another.

Folgmann acknowledges that models will continue to improve, but a company’s clarity on who decides what will not improve on its own. He believes Gartner’s cancellation prediction reflects exactly this point: most failed agent projects are not because the models are inadequate, but because no one defined value or controlled risk—both of which are management issues.

His advice is to treat each agent like a new employee in a regulated role. Write down its rules, limit its scope, and assign a manager to check its work and discuss its performance. Companies that do this will be able to give agents more responsibility over time because they have evidence to justify doing so. Companies that skip this step will spend 2027 explaining why their software did something nobody approved.

It is worth noting that Folgmann’s framework rests on a specific premise: the organization possesses the discipline to execute these management actions. The Gartner prediction he cites—that over 40% of agent projects will be canceled—implies precisely that many companies lack this discipline. His advice has direct operational value for teams that already possess basic operational governance capabilities, but for organizations unwilling to even write rule documentation, a correct diagnosis does not necessarily equate to executability.