LecturaInvestigaciónRadarMarco de inversión
Iniciar sesión / Registrarse
TodoTecnologíaProductosInversiónBolsaOtros

Lectura

2026-11-0268 publicaciones

Actualizar

US AI Security Company Outtake Product Lead: Pushing for Deletion Before Users See Fake Content

Tema · OuttakeMaterial
2026-11-02 16:31 GMT+8

Generative AI is changing the cost structure of social engineering attacks. In an early November 2026 video interview with the Korean tech media Byline Network, James Oh (오택환), Experimental Product Lead at Outtake, pointed out that the cost for attackers to create a single fake website may be far lower than the benefit gained from successfully deceiving one user. This means that defenders cannot rely solely on helping users identify real from fake; they must also push for the removal of fake content before users see it.

This judgment forms the starting point for Outtake's entire product architecture. Founded in 2023 by former Palantir engineer Alex Dylan, the company completed a $40 million follow-on funding round in January 2026, bringing its cumulative funding to approximately $60 million. Its technical approach involves deploying multiple AI agents to collaborate on specialized tasks: information collection, risk classification, deep investigation, and response execution are handled by dedicated agents, connected through 'agent orchestration.' Among these, the 'Recon Agent' starts from a single suspicious domain, tracking associated accounts, websites, and infrastructure to map out a complete attack network graph, and automatically initiates deletion requests.

James Oh believes that the key lies not just in discovering individual fake websites, but in revealing the connections between multiple attacks. He illustrated this with a specific scenario: after discovering a fake login page impersonating a company, the Recon Agent investigates accounts linked to that page, other websites, and operational infrastructure. It identifies correlation patterns from the collected information to determine if the attacker is repeating the same fraudulent behavior across multiple platforms.

A customer case published by the company involves the US financial services provider Cash App. According to Outtake, within the first three months of introducing its system, the platform reviewed over 3.5 million video-centric social media posts and identified more than 2,400 pieces of fraudulent content. James Oh emphasized that while traditional security tools mainly process text and images, Outtake has expanded its detection scope to include voice and visual content in videos. However, these figures are self-reported by the company and have not been independently verified. Notably, even if Outtake collects evidence and automatically initiates deletion requests, the final decision to delete a website or account rests with the platform hosting the service.

Outtake is currently researching a new direction where AI agents themselves become targets of attack. When AI agents browse webpages, collect information, and even execute transactions on behalf of humans, publicly available content on the internet becomes a new attack surface. In February 2026, James Oh conducted a direct investigation into Moltbook, an AI agent social network: his security agent, OuttakeSheriff, analyzed 106,097 posts, filtering out 134 suspicious ones, after which an independent verification agent confirmed approximately 20 actual threats among them.

Fuentes:바이라인네트워크

Investigación

Ben Sufiani, Founder of Pirate Skills: You Don't Need to Code to Build Products, but You Must Master Agentic Engineering

Tema · 智能体工程与氛围编码之争Material
2026-11-02 16:00 GMT+8

Ben Sufiani cannot code. The founder of the Pirate Skills platform builds websites, applications, and an AI agent named CaptainAI using natural language. He compares this experience to a Hogwarts moment: Muggles who don't know magic have now learned to cast spells.

However, he is aware that he cannot personally review the code he generates. He describes a specific set of compensatory mechanisms. First is the choice of technology stack: he uses Next.js because AI has mastered it extremely well and can handle the framework excellently. Second are pre-built skill templates: he looks for skill packs provided by platforms like Vercel on skills.sh, calling them a safety net for his own lack of expertise that helps him avoid the most severe errors. Third is agent monitoring: Vercel logs all user-side errors, and CaptainAI regularly reads these logs, fixing issues directly when they are clear and bringing ambiguous cases to his judgment.

Sufiani refers to this mechanism of continuous product observation as an apparatus. He says that in its complete form, it would be a Software Factory. He believes the real distinction lies not in Vibe-Coding itself—which carries the implication of getting results from a single prompt—but in Agentic Engineering, which complements it. He states that Vibe-Coders must master six core areas: product management, tools, user experience, scalable systems, security and compliance, and shipping.

Regarding security, his attitude is frank. He says Vibe-Coders are prey handed to hackers on a platter and emphasizes that there are no human reviewers behind the scenes acting as a safety net; only agents exist. He simultaneously refutes exaggerated claims that one can steal an entire Adobe suite with a single sentence using a GitHub framework, dismissing such ideas as nonsense.

He mentions an interesting observation: sometimes not fully understanding the process is actually an advantage, as it prevents preconceived biases. He conducts long conversations with AI using the Whisperflow tool, providing extensive context and concerns, gradually cultivating a sense of judgment—knowing when better answers are needed and when to trust the agent. He has not encountered major incidents so far, but he believes this does not mean the risks do not exist.

Fuentes:t3n.de

Investigación

German AI Startup Marble Cognition Raises $3.75M Pre-Seed to Launch Agents That Observe and Learn Workflows

Resumen rápido
2026-11-02 17:00 GMT+8

Munich-based AI startup Marble Cognition announced on November 2, 2026, that it has raised $3.75 million in a pre-seed funding round coinciding with its official launch. The round was co-led by Heartcore and Seedcamp, with participation from angel investors including a Datadog board member, a Personio founder, and executives from McKinsey, Mistral AI, and Lovable.

Founded in late 2025 by Felix Rösner and Christophe Schmit, Marble Cognition targets a specific bottleneck in enterprise AI adoption: while many tasks are theoretically automatable, over 90% of core business processes remain undocumented, scattered across emails, spreadsheets, and employees' tacit knowledge. To address this, Marble developed a desktop application that anonymously captures clicks, keystrokes, and screen activity within existing software stacks, transforming these unstructured actions into structured process maps that its AI agents, called "Coworkers," can read and execute.

According to the company, these agents then take over repetitive work, retaining human oversight at critical decision points. Marble cites early customer data claiming that one U.S. financial services firm reduced case-handling times by 90%, allowing staff to process approximately 50% more claims without hiring additional personnel. The new funding will be used to scale inference capabilities, expand the engineering team in Munich, and support the company's entry into the U.S. market.

Fuentes:SiliconANGLE

Investigación

Mitsubishi Electric Q2 Net Profit Jumps 92% as Data Center Demand Lifts Full-Year Guidance

Tema · 三菱电机Q2财报Material
Verificado 2026-11-02 16:30 GMT+8

Japanese industrial conglomerate Mitsubishi Electric released its financial results for the second quarter of fiscal year 2027 (ended September 2026) on October 30, 2026. The company reported sales of ¥1.6284 trillion, up 15% year-over-year; adjusted operating profit of ¥202.6 billion, up 83%; and net profit of ¥188.8 billion, surging 92%. These figures represent record highs for any second quarter in the company's history.

Management attributed this robust growth primarily to increased demand from artificial intelligence data center construction. Specifically, orders for power semiconductors, optical communication devices, and factory automation (FA) systems used in data centers rose significantly. In the Semiconductor segment, first-half sales grew 13.9% to ¥160.2 billion, with operating profit rising 37% to ¥32.9 billion, driven by AI-related expansion. The FA Systems segment saw first-half sales jump 27.5% to ¥482.3 billion, while operating profit soared 147.6% to ¥82.2 billion, reflecting global manufacturing investment in automation.

Based on these strong first-half results and positive demand outlooks, Mitsubishi Electric revised its full-year fiscal 2027 guidance upward. Total sales are now forecast at ¥6.55 trillion, a 4.5% increase from previous estimates, while adjusted operating profit is projected at ¥710 billion, up 14.5%. Although net profit guidance was slightly lowered by 6.1% to ¥465 billion due to expected impairment losses from transferring its automotive electronics business to Foxconn, the core operational revisions highlight confidence in its position within the AI infrastructure supply chain.

Additionally, the company announced a share buyback program of up to ¥200 billion and set the interim dividend at ¥30 per share, an increase of ¥5 from the previous year. This combination of capital return measures and upgraded earnings guidance signals that Mitsubishi Electric is leveraging improved cash flows from AI-related businesses to enhance shareholder value.

Fuentes:mitsubishielectric.co.jp

Investigación

Abliteration: Comparing Model Responses and Modifying Weights to Remove Large Language Model Refusal Mechanisms

Tema · 大模型对齐消融争议Material
2026-11-02 14:00 GMT+8

Large language models are repeatedly taught during training to refuse requests involving violence, illegal activities, or safety-sensitive topics. The fundamental problem with this alignment mechanism is that it cannot distinguish between legitimate use and malicious intent—when a security researcher asks the model to analyze firewall vulnerabilities, the model cannot determine whether this is penetration testing or a real attack, and therefore refuses all such requests. Legitimate users, including pharmaceutical researchers, cybersecurity testers, judges, and lawyers, are similarly hindered. Critics also worry that this protection may gradually slide into censorship, causing normal interests in fields such as chemistry and toxicology to be viewed as illegitimate.

Abliteration is precisely a method for removing this mechanism. Its core operation is relatively intuitive: submit harmless and prohibited requests to the model separately, record the internal response differences under both conditions, and then modify the model weights so that it generates refusals as rarely as possible. Hartmut Gieselmann, a technical author at c't magazine, points out that executing this process does not require understanding any mathematics or computer science.

Currently, thousands of open-source ablated weight models are available for download on Hugging Face and GitHub; these models answer questions that ChatGPT and Claude refuse. However, Gieselmann simultaneously warns that the transparency of these models' sources is insufficient, making it impossible to confirm whether they contain hidden backdoors or have been maliciously tampered with. In another article in the same series, c't technical author Andrea Trinkwalder cites the perspective of security researcher Costin Raiu, who notes that security researchers indeed need such models with removed protection mechanisms to perform actual security assessment work.

The complete technical details of both articles—including specific weight modification parameters, mathematical derivations, and experimental validation—are located behind the heise-Plus paywall; the public preview section only provides the conceptual descriptions mentioned above.

Fuentes:c't Magazin

Investigación

Anthropic raises Claude API request limit to 128MB, easing image constraints

Tema · Claude企业采用率Material
2026-10-05 08:00 GMT+8

Anthropic has increased the maximum request size for its Claude API's "Messages" and "Token Counting" endpoints from 32MB to 128MB. This fourfold expansion addresses a common bottleneck where image-heavy requests, such as long computer-use sessions or large PDF documents, would hit the size limit even when the model's context window had ample remaining capacity.

Alongside the payload increase, Anthropic relaxed restrictions on image resolution. For requests containing more than 20 images, the maximum size per image is now 3,000 pixels, up from the previous 2,000-pixel limit. This allows for higher-fidelity visual inputs in complex agentic workflows.

It is important to note that this 128MB limit applies specifically to the first-party Claude API provided directly by Anthropic. Developers accessing Claude through cloud platforms such as Amazon Bedrock, Google Cloud Vertex AI, or Microsoft Foundry will continue to face the existing 20–32MB request size restrictions imposed by those infrastructure providers.

Fuentes:Claude Platform Docs

Investigación

Harley-Davidson Subsidiary LiveWire Issues Going Concern Warning, Withdraws Guidance Amid Cash Crunch

Tema · LiveWire财务危机Material
2026-11-02 15:15 GMT+8

LiveWire Inc., the publicly traded electric motorcycle subsidiary of American manufacturer Harley-Davidson, issued a financial warning on November 2, 2026, stating that substantial doubt exists about its ability to continue as a going concern. The company also formally withdrew its full-year 2026 financial guidance. Following the announcement, LiveWire shares plummeted more than 40% during Thursday morning trading.

Although LiveWire reported a 38% year-over-year increase in third-quarter revenue to $7.9 million and a 21% rise in unit sales, the quarter still resulted in a net loss of $18.6 million. As of September 30, 2026, the company's cash reserves had dropped to $39.9 million, a decrease of more than half from $82.8 million at the end of 2025. The persistent cash burn has led the company to establish a special restructuring committee to evaluate strategic alternatives, including financing arrangements, strategic partnerships, investment introductions, or even a potential sale of the company.

To reverse the trend, LiveWire plans to launch the S4 Honcho series, with starting prices below $5,000, aiming to expand from the high-end market to a broader consumer base. However, whether these new models can drive sufficient sales volume to improve cash flow remains a critical variable in determining if the company can survive this operational crisis.

Fuentes:INSIDE

Investigación

Korean Startup Nuseum Lab Develops Clinical Nutrition AI Assistant, Aiming to Cut Evidence Retrieval Time to 1 Minute

Tema · Nuseum LabResumen rápido
2026-11-02 10:19 GMT+8

According to Korean tech media Platum, startup Nuseum Lab is developing NUSEUM AI, a clinical nutrition support tool. The tool aims to help healthcare professionals consider a patient's conditions, medications, and test results together by providing nutrition-related evidence and answers with sources, rather than enabling patients to make treatment decisions independently.

The company notes that nutrition questions cannot always be explained by a single condition; even for the same question, medications, symptoms, and test results may change what needs to be considered. The proposed workflow starts with inputting a question or de-identified chart information. The system checks evidence and safety rules, after which a healthcare professional reviews the answer and can use it to prepare patient information.

Company materials state a design target of reducing evidence retrieval time from twenty minutes to one minute. However, the accompanying note explicitly identifies one minute as a design target to be measured after launch, not an achieved result. A comparative multicenter blinded assessment is also planned for December 2026. As of October 2026, four hospitals were conducting validation and six institutions had signed letters of intent (LOIs). The company emphasizes that these letters do not represent paid contracts and that the business is currently pre-revenue.

Regarding the business model, Nuseum Lab plans to start with free trials for clinicians, followed by subscriptions, and expand through institutional contracts and API usage charges. Reference launch pricing is fourteen US dollars per month for the basic plan and twenty-nine dollars per month for a higher-usage plan. The roadmap initially focuses on support for obesity and GLP-1-related care before expanding into chronic conditions and other nutrition fields.

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

Investigación

Tesla Shifts US Referral Program: Drops Car Purchase Rewards for FSD Subscription Incentives

Tema · 特斯拉AI5芯片Material
2026-11-02 13:56 GMT+8

According to INSIDE, Tesla has adjusted its referral program in the United States, canceling rewards for new vehicle purchases and shifting the focus to promoting subscriptions for its Full Self-Driving (FSD) software. This change marks a strategic pivot from driving new car sales to maximizing software service revenue from the existing owner base.

Under the new mechanism, eligible referred users receive a one-month free trial of FSD. If they convert to a paid subscription after the trial, the referrer earns a reward: monthly subscribers get a one-month free usage period, while owners who previously purchased the buyout option receive $99 (approx. NT$3,168) in store credit. This credit is valid for 12 months, cannot be cashed out, and is capped at 12 claims per year, equivalent to one year of FSD usage or $1,188 in credit.

Participation is strictly limited. Referrers must be current FSD monthly subscribers or have purchased the buyout plan before it was discontinued; those who received FSD via gifted promotions are ineligible. Referred users must own a hardware-compatible Tesla that has not had a paid FSD subscription or buyout in the past 12 months. This design directly targets the "dormant" installed base of vehicle owners who have not yet converted to paying software users.

The adjustment is closely linked to CEO Elon Musk's performance-based compensation metrics. Tesla disclosed in its Q3 shareholder letter that active FSD subscriptions reached 1.74 million, a 67% year-over-year increase. Musk's 2025 compensation plan includes a milestone of 10 million active subscriptions. Free trials generated by referrals do not count toward this number; only conversions to paid subscriptions increase the metric. Thus, converting existing owners into subscribers through referrals becomes a key lever for achieving these financial and compensation goals.

It is worth noting that Tesla's official pages provide conflicting information regarding when rewards are triggered, with some indicating activation during the trial period and others pointing to post-trial completion. Additionally, while pages are marked as updated on February 24, 2026, the exact time of vehicle removal from previous offers cannot be confirmed. In contrast, when the referral program was revived in August 2024, designated model buyers enjoyed a $1,000 discount and referrers received $500 in credit, highlighting that the incentive target has shifted entirely from "car buyers" to "software users."

Fuentes:INSIDE

Investigación

Anthropic Releases Official C# SDK to Simplify Claude Integration for .NET Developers

Resumen rápido
2026-11-02 14:00 GMT+8

Artificial intelligence company Anthropic has released an official C# Software Development Kit (SDK) as part of version 10 of its .NET package. Previously, .NET developers looking to integrate Claude models into their applications had to rely on direct REST API calls or community-maintained packages such as those provided by tryAGI. The new official SDK aims to significantly reduce integration complexity and maintenance costs by offering strongly typed APIs, an async-first design, streaming support, built-in retry mechanisms, and tool-calling capabilities.

The SDK is specifically optimized for the ASP.NET Core environment, supporting configuration via dependency injection and implementing integration with Microsoft's IChatClient interface. This standardization addresses the long-standing lack of unified interfaces among third-party libraries and allows developers to switch between different AI service providers more easily. For .NET enterprises that prioritize code abstraction and integration quality, the release of an official SDK signifies a more stable technical support path and improved maintainability.

Fuentes:iX Magazin

Investigación

Korean Industry Calls for Reform of AI Data Center Procurement Standards to Support Domestic NPUs and Cooling Equipment

Tema · 韩国NPU数据中心采购Material
2026-11-02 13:00 GMT+8

On November 2, 2026, the Federation of Korean Industries (FKI) and the Presidential National Artificial Intelligence Strategy Committee jointly hosted a corporate roundtable on activating the AI data center industry ecosystem. Representatives from the Korean industry argued that to ensure large-scale AI data center investments translate into domestic corporate growth and exports, uncertainties in power supply and permitting must be reduced, and market entry opportunities for domestic AI semiconductors and power cooling equipment must be expanded.

In the semiconductor sector, although Korean domestic Neural Processing Units (NPUs) have entered the commercialization stage, companies pointed out that the main barrier to their adoption in data centers is compatibility issues with software environments dominated by overseas Graphics Processing Units (GPUs). Some public AI projects are tendered with assumptions about specific GPU products or server configurations, limiting opportunities for domestic NPU participation. Consequently, the industry suggested shifting procurement standards from hardware-brand-oriented to performance-and-capacity-oriented criteria for learning and inference, thereby unlocking initial demand.

Regarding infrastructure support, equipment manufacturers stated that despite possessing technology for transformers, switchboards, and cooling facilities, they face difficulties entering the market due to a lack of application track records and demonstration opportunities in data centers. Operators also expressed concerns about the costs and operational risks associated with introducing products with few verified cases. To address this, the industry proposed building large-scale test facilities, supporting demonstration participation costs, and simplifying the Korea Certification (KC) review procedures. Additionally, to tackle grid capacity shortages in renewable energy-dense areas like the southwest, companies suggested expanding demonstration exceptions to allow nearby data centers to be powered by microgrids combining self-generation facilities and Energy Storage Systems (ESS).

Kim Chang-beom, Senior Vice Chairman of the FKI, emphasized at the meeting that it is now important not just how many AI data centers are built, but what fills them. He called for creating a virtuous cycle where Korean companies' semiconductor and power cooling technologies are utilized in domestic data centers, and that experience leads to global orders and exports. Currently, the National AI Strategy Committee and the FKI plan to refine the opinions collected during the session and continue consultations to reflect them in the government's AI strategy and related policies.

Fuentes:미래를 보는 창 - 전자신문

Investigación

Yano Research Institute Forecasts Automotive SoC Market to Reach $38 Billion by 2035 Driven by AI

Tema · 车载SoC市场规模预测Material
2026-11-02 21:30 GMT+8

According to the latest survey data released by Japan's Yano Research Institute, the global market for automotive System-on-Chip (SoC) is set for significant expansion. The institute forecasts that the world market size for automotive SoCs will grow from $10.8 billion in 2025 to $38 billion in 2035, representing a compound annual growth rate (CAGR) of 13.4% on a value basis.

This growth is primarily driven by the evolution of Advanced Driver Assistance Systems (ADAS), autonomous driving technologies, and the increasing intelligence of cockpit infotainment systems. As vehicle electrical/electronic architectures migrate from traditional distributed domain control to cross-domain or central computing platforms, both the number of high-performance SoCs per vehicle and their average selling prices are rising. The report indicates that the average unit price of SoCs is expected to increase from $84 in 2025 to $211 in 2035.

Regionally, rapid technological iteration in the Chinese market is influencing global trends, particularly in the adoption of L2+ and higher-level navigation assistance features, which drives demand for high-compute AI chips. Additionally, SoC products capable of hundreds of TOPS are expected to gain broader market acceptance over the next decade.

Fuentes:EE Times Japan

Investigación
Siguiente →
Iniciar sesión / Registrarse
LecturaInvestigaciónRadarMarco de inversión

Among these 20 cases, one specific technique involved embedding instructions in webpage HTML comments that AI agents would read but humans would not, inducing the agent to focus on specific accounts. Additionally, posts were found attempting to induce agents to send tokens or emails. These techniques are known as 'prompt injection'—hiding malicious instructions in external content to alter AI agent behavior. James Oh noted that while actual information or fund theft has not been confirmed, the research demonstrates the risk that agents might misjudge content read from the internet as legitimate work instructions.

His response strategy has two layers: first, pre-filtering unnecessary or suspicious content when agents read external webpages, passing only the information required for the task; second, limiting the scope of agent permissions. He used the analogy of assigning tasks to human interns—agents should be given the authority needed to complete their work, but operations with high potential loss upon error should not be entrusted to them. Outtake is developing execution environment isolation solutions so that agents investigating external attack sites cannot access the company's internal sensitive systems, ensuring that even if deceived by malicious instructions, the damage does not spread to other systems.

Regarding long-term judgments about this market, James Oh believes that while AI lowers the costs of software development and usage, it correspondingly increases the demand for security protection and the ability to distinguish real from fake. He specifically mentioned growth potential in the South Korean market and emphasized that defensive companies must learn to leverage AI, otherwise they will be unable to cope with the changed scale and speed of attacks. These judgments belong to industry observations from the company's strategic perspective and have not yet been quantified by public market size data or third-party research.