ReadingResearchRadarInvestment framework
Sign in / Sign up
AllTechnologyProductsInvestingPublic marketsOther

Reading

2026-11-19129 posts

Refresh

Chinese AI Firm DeepSeek Adjusts API Pricing: In-Model Search Now Token-Billed Only

Topic · DeepSeek API PricingMaterial
2026-11-19 16:51 GMT+8

Chinese AI company DeepSeek adjusted its API billing rules effective 12:00 Beijing time on November 18. According to a report by Zhidx, when developers execute tasks using DeepSeek models that trigger server-side search or web scraping, these operations will no longer incur separate per-call fees. Instead, they are included in the standard token usage billing. This significantly reduces the marginal cost for developers building AI Agents that require frequent online information retrieval.

It is important to note that this adjustment does not make all search and scraping interfaces free. If developers bypass model calls and directly request the independent Search API or Fetch API, they must still pay based on the number of successful requests. The official pricing page lists both services at 5 yuan per 1,000 successful requests. This distinction aims to encourage developers to integrate search capabilities naturally within model inference rather than using DeepSeek solely as an external search engine.

Accompanying this billing change, DeepSeek Harness released version v0.2.2-rc.5. According to the GitHub changelog, when users employ DeepSeek models within this framework, Harness defaults to DeepSeek's server-side search to benefit from the waived per-call fees. However, when users employ other models, Harness defaults to calling the independent DeepSeek Search API, where search costs are still charged according to interface rules.

This cost optimization for developers follows a tightening of consumer-side experience. On November 16, some users of the DeepSeek app and web interface encountered issues such as interrupted long conversations, reduced input limits, and daily quota restrictions, with system prompts citing "insufficient computing power." Many users called for paid services on social media to retain their previous experience. The contrast between resource throttling for general users and tool-call cost reductions for developers reflects DeepSeek's attempt to allocate limited computing resources toward the high-value developer ecosystem, aiming to attract more enterprises and individuals to build applications on its models by lowering barriers.

Sources:zhidx.com

Research

California Proposes Emergency Shutdown Powers for High-Risk AI Models in New Safety Report

Topic · California AI Emergency Shutdown RegulationMaterial
2026-11-19 15:52 GMT+8

The California state government released a policy report titled "AI Safety in 2026 and Beyond" on November 18, 2026, outlining legislative recommendations to strengthen artificial intelligence oversight. The proposal grants the state government emergency powers to directly order companies to halt high-risk AI models under specific conditions, without requiring a prior court injunction. This marks a significant step toward more coercive local regulation of potential AI out-of-control scenarios.

The report mandates that large advanced AI developers prepare emergency response measures, including stopping AI agent operations, revoking system access, and isolating computing resources. If a model poses an immediate risk of catastrophic accident or severe loss of control, and the company fails or refuses to address the danger, the state can intervene to force a shutdown. Additionally, the plan suggests establishing continuous third-party safety verification, requiring independent assessment bodies to monitor training, development, and deployment phases and submit full results to the government.

Addressing the growing concern over Recursive Self-Improvement (RSI)—where AI systems autonomously develop stronger successor models without human intervention—California proposes specific restrictions. The report recommends requiring companies to disclose the extent of AI involvement in research and the computational resources dedicated to automated development. If AI-driven progress exceeds current safety measures and evaluation capabilities, posing a substantial risk of catastrophic harm, the state would have the authority to limit the proportion of compute resources allocated to such automated R&D.

Regarding incident reporting, the report suggests expanding the definition of major AI safety events to include model failures and attempts to steal model weights, while creating channels for near-miss incidents. The deadline for reporting most major events is proposed to be shortened from the current 15 days to 48 hours, with a 24-hour requirement maintained for events with imminent catastrophic risk.

It is important to note that these measures remain at the policy recommendation stage and require passage through the California legislative process to become law. Previously, California enacted SB 53 in 2025 and signed SB 813 and AB 1405 in 2026, establishing certification for independent verification organizations and registration for AI auditors. This report follows Governor Gavin Newsom's September directive to study emergency shutdown mechanisms, presenting a concrete legislative blueprint as regulators seek a new balance between fostering innovation and preventing systemic risks.

Sources:iThome

Research

Why AI Still Cannot Autonomously Adjust Process Parameters in Semiconductor Manufacturing

Topic · AI Autonomy in Semiconductor ManufacturingMaterial
2026-11-19 16:01 GMT+8

At the SEMICON West 2026 conference, a recurring question was discussed: If AI can already identify process solutions invisible to human engineers, why is it still unable to autonomously adjust production parameters?

David Fried, Chief AI Officer at Lam Research, described a specific scenario: The potential recipe space for an etcher involves approximately 10^23 combinations. Constrained by historical experience, human engineers tend to revert to parameter regions where problems were solved ten years ago. In contrast, AI can point to operating regions that engineers have never considered. Vikas Sharma, Senior Director of Pathfinding for R&D at Intel Foundry, provided an even more extreme example: A process layer involving 350 parameters has a theoretical combination count of approximately 10^700. Engineers and equipment suppliers spent a year developing a minimally acceptable recipe, but after introducing AI-assisted optimization, multiple solutions superior to the original baseline were found within just a few months.

The capability has been proven, yet the authority remains ungranted. Where are the obstacles?

The first obstacle is that models identify statistical relationships rather than physical mechanisms. Eric Guichard, Head of AI Manufacturing Business at Silvaco, pointed out that surrogate models are reliable when interpolating within the region represented by their training data, but become less reliable when extrapolating beyond it. The fact that AI identifies new operating regions—precisely what makes it most valuable—also means that engineers have the least experimental evidence on hand to validate its predictions. The relative importance of pressure, temperature, voltage, and gas flow may differ completely across different operating regions; a model that performs well in one region cannot be assumed to hold true in another.

The second obstacle is data fragmentation. Information in semiconductor manufacturing is often scattered across different companies and systems. In advanced packaging, a foundry provides the die, another company produces the substrate, and an OSAT (Outsourced Semiconductor Assembly and Test provider) completes assembly and testing; these three parties typically cannot share the same manufacturing history. It is even more difficult because physical properties such as substrate warpage, thermal stress, and material behavior change continuously during manufacturing. These changes are often not captured by existing measurements, and models relying on nominal material properties may miss them entirely. Kenneth Larsen, Senior Director of Product Management at Synopsys, summarized the core issue: If the training range, calibration source, and approved use are not transmitted along with the model, it may be sufficient for early exploration but unsafe for design closure. What is missing is not more data, but the context that tells the next tool or team how the data should be used.

The third obstacle is the time lag between model updates and production. Sharma explicitly stated that the industry is far from continuous learning; the actual practice is updating models quarterly or monthly. A fab may process 10,000 wafers between two updates, and the data generated by these wafers is not rapidly used to improve the model or agent learning. More perplexingly, if the model adjusts equipment settings between updates, engineers need to know exactly what changed and whether the updated behavior remains within verified limits. When investigating failures that emerge weeks later, it may be necessary to reconstruct the state of the model when it processed specific batches.

Sources:semiengineering.com

Research

US Industrial Automation Giant Rockwell Automation Signs Deal to Acquire Austrian AMR Maker AGILOX

Topic · Rockwell-AGILOX AcquisitionMaterial
2026-11-18

US-based industrial automation company Rockwell Automation (NYSE: ROK) announced on November 18, 2026, that it has signed a definitive agreement to acquire AGILOX, an Austria-based manufacturer of autonomous mobile robots (AMRs). The transaction is subject to regulatory approval and is expected to close in the second quarter of Rockwell’s fiscal year 2027.

Founded in 2017, AGILOX specializes in high-end AMRs driven by self-developed artificial intelligence and swarm networking technology, with over 2,500 forklift AMRs deployed globally. Rockwell stated that upon completion, AGILOX’s proven forklift AMR portfolio will complement OTTO by Rockwell Automation’s existing flatbed AMR platform to create a comprehensive autonomous material movement offering. This integration elevates AGILOX from a best-in-class point solution to a cornerstone of a world-class production logistics capability.

Strategically, the acquisition combines OTTO’s strong presence in North America with AGILOX’s established footprint across Europe. AGILOX was previously backed by private equity firm Carlyle, which made a majority growth investment in 2021, during which time AGILOX quadrupled its revenue. Rockwell noted that the deal will more than double its AMR business and provide AGILOX access to Rockwell’s deep customer relationships in strategic verticals such as consumer packaged goods and life sciences.

Blake Moret, Chairman and CEO of Rockwell Automation, said that combining the technologies creates a best-in-class platform setting a new standard for autonomous material movement. Upon closing, AGILOX will report to Rockwell’s Intelligent Devices operating segment.

Sources:Rockwell Automation

Research

Maum AI’s Son Byung-hee: Standardizing Robot Behavior Data Formats Is Essential for Cross-Robot Experience Sharing

Topic · Robotics Data StandardizationMaterial
2026-11-19 10:25 GMT+8

Large language models can learn from the text and images already available on the internet, but data regarding robots grasping objects, operating equipment, or avoiding obstacles is scarce online. Son Byung-hee, Head of Defense and Robotics at the South Korean AI company Maum AI, points out that robots must collect this data using their own bodies, and that collection is merely the first step.

Son Byung-hee believes that while bottlenecks exist at every stage of the physical AI learning loop, a critical hurdle lies in data aggregation: because different robots record data in disparate formats, information from various manufacturers is difficult to merge into a single training set. As Chairman of the Standards Council of the Korea Physical AI Association, he is collaborating with the Electronics and Telecommunications Research Institute (ETRI) to discuss standardization benchmarks for robot behavior data.

His solution centers on a core data structure called the 'episode.' An episode records the complete experience of a task from start to finish, comprising five elements: Task (what needs to be done), Observation (what was seen), State (the conditions of the robot and environment at that moment), Action (what operation was executed), and Outcome (whether it succeeded or failed, and at which stage). Son Byung-hee emphasizes that simply labeling success or failure is insufficient. He cites the example of a robot picking up a cup: whether it gripped it stably, succeeded by chance, or recovered after slipping mid-air carries vastly different implications for training. Therefore, episodes must link context, action, and outcome, enabling AI to learn causal relationships beyond mere action sequences.

The goal of standardization is to achieve cross-embodiment: allowing experiences gained by one robot morphology to be utilized by another. Son Byung-hee explains that even if end-effectors differ significantly—such as a two-fingered gripper versus a five-fingered human hand—the operational knowledge regarding what object was picked up, from where, and placed where can be shared, provided these experiences are recorded in a common format. He argues that data standardization will make the realization of cross-embodiment learning significantly easier.

Maum AI has established a Physical AI Data Factory at its headquarters in Pangyo, South Korea, utilizing NVIDIA Isaac Sim simulation software and ALOHA series bimanual teleoperation devices to collect data. Son Byung-hee describes a closed-loop process: actual robots collect data → train Vision-Language-Action (VLA) models → conduct experiments with varied conditions in simulation environments → deploy models back to real robots for validation → re-collect successful and failed results as new data. He notes that bottlenecks exist at every stage of this cycle, with data format inconsistency causing aggregation difficulties being the specific issue he highlights.

Son Byung-hee also acknowledges that standardization cannot solve all problems. A sim-to-real gap exists between simulation and reality; no matter how refined the simulation, real-world robot data remains essential. The Korea Physical AI Association is currently participating in planning for the Ministry of Land, Infrastructure and Transport’s Urban Environment Physical AI Virtual Learning Field, aiming to connect real urban spaces with virtual environments to provide a verification foundation for robot mobility and operational learning.

Sources:바이라인네트워크

Research

Singaporean Venture Creator Origgin Ventures and Global Business Builder KILSA Global Sign MOU to Launch 'Co-Creation' Model in Korea

Quick take
2026-11-19 13:47 GMT+8

Singapore-based 'Venture Creator' Origgin Ventures and global business builder KILSA Global signed a memorandum of understanding on November 18 in Seongnam, South Korea, to introduce the 'Venture Co-Creation' (VCC) model to the Korean market. Unlike traditional external equity investment, this model involves actively identifying technology from universities and research institutes, co-founding companies with researchers, and providing pre-seed funding along with management talent.

Under the agreement, the two companies plan to make pre-seed investments in 3 to 5 early-stage Korean startups each year, focusing on sectors where Korea has strong R&D capabilities, such as semiconductors, secondary batteries, robotics, and biotech. Clarence Tan, founder of Origgin Ventures, stated that the firm would apply its nine years of VCC experience to help validate the commercial viability of Korean technologies and directly participate in startup formation. KILSA Global will contribute its 'Global Business Building' (GBB) model, which integrates overseas demand and local partners from the inception stage, particularly targeting markets in Singapore and ASEAN.

The partnership aims to address the gap between Korea's robust research output and its commercialization infrastructure. The companies noted that while Korea lacks sufficient commercialization capital and sales channels, Singapore serves as an ASEAN hub for global capital despite its smaller domestic market. Additionally, they plan to collaborate with Korean accelerator Infobank iExcel to facilitate follow-on investments and are considering establishing a joint venture after a one-year pilot period.

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

Research

LG Electronics Restructures for AI: New Data Center Cooling Division, Promotions Up 33%

Topic · LG Electronics Data Center Cooling RestructuringMaterial
2026-11-19 15:09 GMT+8

South Korean electronics giant LG Electronics announced its regular personnel and organizational restructuring for 2027 on November 19, with "robotics" and "data centers" as key themes. The company has made AI data center cooling solutions and robotics central to its strategy by establishing a new "Applied Business Division" under the ES (Enterprise Solutions) unit, dedicated to addressing data center cooling demand and expanding orders. Led by Vice President Lee Jae-hyun, formerly head of the Global Operation Center, this division will be supported by a newly established production entity in Virginia, USA, aimed at increasing global chiller capacity.

In terms of personnel, LG Electronics significantly expanded its executive team to support these new growth areas. Total promotions reached 44, an increase of approximately 33% from last year's 33. Notably, Baek Seung-tae, Head of the HS (Home Entertainment/Appliances) Unit, and Lee Sang-woo, Head of the MS (Media Solutions) Unit and Display Business, were promoted to President. Additionally, LG Electronics moved overseas regional representatives for North America, Europe, and other regions directly under the CEO, bypassing the Overseas Sales Headquarters to reduce decision-making layers and improve responsiveness to global market changes. These moves signal LG's intent to accelerate the conversion of AI infrastructure technologies into actual revenue through internal structural realignment.

Sources:미래를 보는 창 - 전자신문

Research

Radio Free Mobile Founder Richard Windsor: Anthropic Must Win the Model Race, While Hardware Companies Only Need AI Demand to Exist

Material
2026-11-19 13:50 GMT+8

Anthropic completed its IPO with a share price of $83, selling 980 million shares and raising $81 billion, upscaling the original deal in both price and volume. Richard Windsor views this as a positive signal in itself—it indicates that despite the broader sentiment on AI trades being hammered by Nvidia’s slump, rising interest rates, and the US election, there remains appetite for risk in AI-related investments. However, his core argument does not rest on the pricing outcome of this IPO, but rather on the fundamental difference in risk structures between Anthropic and companies such as Nvidia, Broadcom, and Qualcomm.

Windsor’s argument begins with a simple distinction: all picks-and-shovels companies—Nvidia, Broadcom, Marvel, Qualcomm, MediaTek, Arm—benefit merely from the existence of AI demand, regardless of which model company ultimately captures it. Every market signal suggests that demand exists and continues to grow. Anthropic is different: it relies on two conditions simultaneously—demand growth, and the realization that its models become the operating system of AI rather than those of OpenAI, open-source projects, or any other competitor. Furthermore, it must be able to charge service prices high enough to cover compute costs, which RFM Research consistently estimates are significantly higher than OpenAI’s.

This dual dependency creates a specific failure scenario: if OpenAI’s products are at least as good as Anthropic’s (Windsor notes that GPT-6 has materially improved), OpenAI could undercut prices and trigger a price war. Windsor believes this will not happen immediately because Anthropic’s current numbers are healthy. He provided RFM Research’s cash flow framework: although Anthropic has committed to spending nearly $900 billion on compute, this is spread over several years, meaning it will spend around $100 billion in 2027 and perhaps $150 billion in 2028. RFM estimates Anthropic will generate approximately $120 billion in revenue in 2027 and $200 billion in 2028. If these targets are met, Anthropic will not have much trouble covering its compute bills from revenue, which is why Windsor is not overly concerned about the $900 billion commitment.

However, Windsor explicitly defines the boundaries of this conclusion: it depends entirely on Anthropic’s ability to maintain its competitive edge. Once embroiled in a price war, revenues will suffer a double whammy from both market share loss and gross margin declines, causing cash flow to deteriorate rapidly. This would push Anthropic deeper into losses, forcing it back to the market for more capital. Windsor estimates that with the $81 billion raised via IPO plus existing funds, Anthropic holds roughly $200 billion in cash, allowing it to sustain compute spending for at least two years even in a worst-case scenario. But after two years, if the competitive landscape has not improved, the situation becomes critical.

Windsor argues that the market is currently ignoring this risk entirely, instead pricing in another scenario: that Anthropic will dominate the global enterprise market, generating 85% gross margins and substantial cash. Based on this judgment, he expects the stock price on the first day of trading to rise 10% to 15% above the IPO price. However, he notes that this risk appetite for Anthropic has not yet spilled over to other companies, including those that will directly benefit from the $900 billion compute commitment. He concludes that Nvidia and Qualcomm—both of which he holds positions in—remain far superior investment choices on a risk-adjusted basis compared to Anthropic, and he chooses to watch Anthropic’s first-day fireworks from the sidelines.

Sources:radiofreemobile.com

Research

AI Firm Anthropic Launches Record IPO: Raises $107.3 Billion at $1.5 Trillion Valuation

Structural
2026-11-19 15:00 GMT+8

Artificial intelligence company Anthropic began trading on public markets on November 19, 2026, completing the largest initial public offering (IPO) in history. The total proceeds from the offering reached $107.3 billion, establishing an initial market capitalization of approximately $1.5 trillion. This figure significantly surpassed the previous record set by Elon Musk's SpaceX earlier this June.

Due to overwhelming investor demand, Anthropic increased the number of shares offered from the originally planned 825 million to 980 million, pricing them at $83 per share, above the initial range of $70 to $78. This portion of the public offering alone raised approximately $81.3 billion. Additionally, Anthropic sold about $26 billion worth of stock directly to strategic investors such as Google, Microsoft, and AMD. According to CNBC, subscription orders for the deal totaled around $500 billion, indicating strong institutional confidence in frontier AI models.

As a primary competitor to ChatGPT developer OpenAI, Anthropic's listing marks the entry of major AI laboratories into the public capital markets. Although the company faces pressure from losses due to massive infrastructure investments, and the industry contends with concerns over AI safety risks and data center construction controversies, the successful pricing demonstrates that investors are willing to bear risk for potential long-term returns. OpenAI plans to follow suit next year, making Anthropic's market performance a key reference point for the entire AI sector's valuation framework.

Sources:t3n.de

Research

The Real Bottleneck for Enterprise AI Agents: How Probabilistic Reasoning Connects to Deterministic Business Rules

Topic · Agentic Enterprise ArchitectureMaterial
2026-11-19 15:07 GMT+8

The most frequently cited obstacle to AI agents entering enterprises is 'insufficient model capability' or 'poor data quality.' The 'Agentic Enterprise Architecture' proposed by Salesforce at its 2026 Dreamforce conference offers a more specific diagnosis: the problem lies between probabilistic reasoning and deterministic execution.

Large language models excel at understanding intent, summarizing content, and making suggestions—all probabilistic operations where the same question might yield different but equally reasonable answers. However, enterprise operations rely on deterministic rules: refunds require checking order records and approval authority, discounts cannot exceed preset limits, and data access is constrained by roles. Between the model determining that a refund 'should be issued' and the actual execution of that refund stands an entire deterministic chain involving order queries, policy matching, permission verification, and approval workflows.

Marc Benioff, Chairman and CEO of Salesforce, stated at Dreamforce 2026: 'You cannot run a business with just an AI model.' This does not deny model capabilities but points out that judgment and execution belong to different systems. A Deloitte survey of 501 managers and executives at U.S. companies piloting AI agents corroborates this disconnect: only 5% believe their business processes are 'very ready' to introduce AI agents, 72% report difficulties connecting data scattered across multiple systems, and 70% cite trust and governance as primary challenges.

Salesforce's proposed solution is a four-layer architecture. The bottom layer is Data 360, which connects and integrates enterprise data scattered across multiple systems, providing a unified data foundation for AI. Above it sits Customer 360, which hosts the business logic and workflows accumulated in existing applications such as sales, service, and marketing. The third layer, Agentforce, serves as the digital workforce that actually executes tasks, handling specific work like customer responses and opportunity discovery. The top layer, AI Force, is the interaction interface between humans and AI agents, calling upon enterprise data and functions through existing work environments like Slack and Claude.

A key design principle of this architecture is that AI does not independently judge everything; instead, it operates within the enterprise's established access permissions, security policies, and approval workflows. Probabilistic reasoning is responsible for understanding intent and proposing solutions, while deterministic systems are responsible for verifying compliance and executing operations. The two are connected through architectural layers, rather than allowing the model to directly manipulate business systems.

From published cases, Anthropic shortened complex transaction processing cycles by 60% and reduced annual operating costs by approximately $4.5 million after using Salesforce Revenue Cloud and Slack; Adecco shortened recruitment preparation time by 93% after connecting over 60 scattered systems to Data 360. However, these figures are self-reported by Salesforce and have not been independently verified.

Sources:바이라인네트워크

Research

Chinese Appliance Giants Enter AI Data Center Cooling Market, Challenging Samsung and LG

Topic · China Appliance AI CoolingMaterial
2026-11-19 15:00 GMT+8

As power density and thermal management requirements for Artificial Intelligence Data Centers (AIDCs) surge, China's three major home appliance manufacturers—Midea Group, Gree Electric, and Haier Smart Home—are aggressively entering this high-growth infrastructure market. According to a report by Electronic Times, these companies are utilizing their existing advantages in compressors, fluid control, and large-scale manufacturing to compete head-on with Samsung Electronics and LG Electronics, who have identified AIDC cooling as a 'next-generation growth engine.'

Midea Group has joined the MGX ecosystem partner list for NVIDIA's next-generation Vera Rubin platform. The company is investing over 1 billion RMB in Guangdong Shunde to build a liquid cooling production base, scheduled to come online in August 2027. Additionally, Midea has supplied magnetic bearing Coolant Distribution Units (CDUs) in bulk to large cloud service providers in Southeast Asia and is collaborating with Singapore's Keppel Corp to target the Asia-Pacific market. In June, Midea also signed a strategic cooperation agreement with Alibaba regarding AIDC liquid cooling.

Gree Electric emphasizes its strength in producing core components like magnetic bearing compressors. In June, Gree launched an integrated CDU product with a maximum capacity of 1,200 kW and secured orders for liquid cooling systems in data centers in Thailand and Malaysia. This month, Gree won orders worth over 300 million RMB for data center and base station HVAC equipment from China Mobile and China Tower.

Haier Smart Home is investing over 1 billion RMB in Rayong, Thailand, to build a production base for magnetic bearing chillers and CDUs, targeting operation in 2027. Reports indicate that Haier supplied 48 magnetic bearing chillers to a large data center in Malaysia, reducing the Power Usage Effectiveness (PUE) from 1.45 to 1.38. Haier's Q3 data center revenue nearly quintupled year-over-year, with a goal of reaching 10 billion RMB in revenue within three years.

Facing the offensive from Chinese firms, South Korean electronics giants have adopted different response strategies. Samsung Electronics acquired Germany's FläktGroup last year for 1.5 billion euros; 25% of FläktGroup's revenue comes from data centers. Samsung has also hired former Schneider Electric executives and is reportedly negotiating the acquisition of US-based Modine Manufacturing, while planning to launch its CDU production line in Gwangju in early 2028.

LG Electronics relies on its own technology to enter the NVIDIA supply chain directly. By the third quarter, LG's cumulative orders in the AIDC cooling sector exceeded 3 trillion KRW, driven largely by a long-term chiller supply agreement with US-based AirControl Concepts for 5 GW of capacity. Three different power-rated CDU products from LG have already received NVIDIA certification, with plans to expand up to the 4 MW level.

Despite intense competition, the three Chinese companies have not yet disclosed specific liquid cooling business revenues, and their mass shipment timelines are concentrated in 2027–2028. This leaves Samsung and LG a window of approximately one to two years to consolidate their positions in the high-end market and refine their localized service networks.

Sources:미래를 보는 창 - 전자신문

Research

Lee Seung-kyung, Head of AhnLab's Artificial Intelligence Development Office: AI Agent Completing a Goal Does Not Equal Safe Execution; Process Monitoring Must Cover the Entire Task Chain

Topic · AhnLab AI Agent SecurityMaterial
2026-11-19 14:31 GMT+8

Has an AI agent completed its task safely? The answer from Lee Seung-kyung, Head of AhnLab's Artificial Intelligence Development Office at South Korean cybersecurity firm AhnLab, is no. In an interview with Byline Network, he proposed a core judgment: a correct result does not prove that the process was safe, and the execution path of an AI agent is precisely the hardest part to predict.

This judgment stems from AhnLab's self-developed AI penetration testing practice. Penetration testing is a security test that simulates an attacker's perspective to check for system vulnerabilities. Lee Seung-kyung discovered that when an AI agent is assigned the same penetration testing target, the tools it chooses, the systems it accesses, and the order in which it executes them differ every time. This means that security teams cannot exhaustively enumerate all possible execution paths in advance, nor can they determine whether boundaries were crossed based solely on the results afterward.

He gave a specific example: suppose an AI agent is allowed to test a specific vulnerability on a certain server. During its exploration, the agent might incidentally access other production systems on the same network. The penetration test itself was successful, but it touched upon unauthorized targets. Lee Seung-kyung believes that in such cases, one cannot deem the execution safe simply because the goal was achieved.

Based on this observation, AhnLab has proposed a control framework that is broader than traditional AI security. Traditional approaches mainly filter inputs and outputs—detecting malicious prompts and intercepting sensitive information leaks. However, between inputs and outputs, an AI agent goes through multi-step planning, tool invocation, and intermediate result processing, each step potentially introducing risks. AhnLab argues that the security boundary must extend to the execution process itself: what permissions the agent used, which systems it accessed, and which tools it invoked all require continuous monitoring.

Specifically, AhnLab recommends a three-layer control. The first layer is pre-event permission division: data querying, data modification, and external transmission must be authorized separately; having the permission to query customer information does not imply default permission to export it. The second layer is runtime monitoring: even if each individual operation falls within the permitted scope, a combination of operations may create risks, requiring real-time detection of deviations from the expected path. The third layer is instant interruption: once dangerous execution is detected, the system must be able to immediately stop the current task rather than waiting until completion to review.

Lee Seung-kyung also specifically highlighted risk propagation in multi-agent collaboration scenarios. If one agent is injected with a malicious command, it may delegate erroneous tasks to other agents, causing risks to spread across multiple systems. Therefore, task handoffs and information exchanges between agents also need to be included within the scope of security audits.

Sources:바이라인네트워크

Research
Next →
Sign in / Sign up
ReadingResearchRadarInvestment framework

These three obstacles collectively point to a conclusion: The autonomy of AI in semiconductor manufacturing is not a single technical problem, but a verification infrastructure problem. Fried admitted that teams spend more time deciding whether to use full physical simulation, surrogate models, machine learning, deep learning, or agentic AI to solve specific problems than they do actually building and implementing these systems. The packaging inspection case presented by Ingu Yin Chang, Executive Vice President of ASE Group, illustrates the current status: AI can identify a thousand shades of gray difference in X-ray images, but the system can only trigger equipment responses based on this once humans physically confirm through traditional manufacturing methods that a certain shade corresponds to a real defect. Between identifying anomalies and autonomously correcting them lies a whole set of verification, governance, and traceability mechanisms that have not yet been built.

Notably, even Anthropic, the developer of Claude, uses Salesforce as its core CRM and Slack as its primary collaboration tool, while Salesforce simultaneously integrates Claude into its AI agents and business interfaces. Model developers and enterprise software companies serve as both clients and partners to each other, illustrating that 'AI model capability' and 'enterprise execution infrastructure' are indeed issues on two different levels.

The practical significance for readers lies in the evaluation framework: when a company claims to have deployed AI agents, one can ask whether it has solved the connection problem from probabilistic reasoning to deterministic execution—whether data is unified, whether business rules are callable by AI, and whether permissions and approvals are integrated into the AI's execution path. If these are missing, AI may be able to provide the correct answer, but cannot execute it safely.

Park Se-jin, Representative of Salesforce Korea, summarized the priority concerns of Korean enterprises: customers first ask not 'how smart it is,' but 'how trustworthy it is.' Based on this, the article concludes that as AI agents further enter enterprises, the competitive focus is expected to shift from model performance or agent quantity to actual business outcomes and controllable operational architectures.

Regarding autonomy grading, AhnLab suggests deciding whether human approval is needed based on the importance of the task, scope of impact, data sensitivity, and reversibility. General information queries can be highly autonomous, but operations involving financial transactions, deletion of core data, or changes to system configurations should require human confirmation. Agents should not expand their own permissions and must pause to request human judgment when encountering unpre-authorized tasks.

It should be noted that AhnLab itself is a security product vendor, and its process security framework is directly related to its commercial product lines. Currently, AhnLab's generative AI security product, SecureBridge, primarily focuses on input and output filtering, while process-level security controls are still in a stage of technological development. Lee Seung-kyung also observed that current enterprise customers' AI security needs remain concentrated on protecting the inputs and outputs of generative AI, and the demand for controlling the execution process has not yet fully manifested.

However, as AI agents move from demonstrations to actual enterprise workflows, the unpredictability of execution paths will become an unavoidable engineering problem. AhnLab's argument provides a useful starting point for thought: the unit of security assessment should not be a single input or output, but a complete chain of task execution.