독서리서치레이더투자 프레임워크
로그인 / 회원가입
전체기술제품투자상장시장기타

독서

2026-10-22168 게시물

새로고침

Open-source AI model gateway LiteLLM hit by supply chain attack; official release pipeline compromised to steal keys

주제 · LiteLLM供应链安全중대
2026-10-22 13:15 GMT+8

The open-source AI model gateway LiteLLM suffered another supply chain attack on October 21, 2026. According to iThome, hackers stole GitHub login credentials from three LiteLLM engineers and used an account with administrative privileges to disable code modification protections. They redirected the official release system's source code reading location to a typosquatting repository with a similar name, allowing malicious code to be built and published through the official pipeline to platforms like PyPI and Docker Hub.

Affected versions include 1.107.0 on PyPI and multiple download options in Docker Hub and GitHub Container Registry repositories. Developers who downloaded images marked as 'latest' at the time may have obtained the malware even if they did not specify version 1.107.0. Customers obtaining enterprise versions from private repositories were unaffected. Security firm SafeDep analyzed the compromised package and found it injects malicious files into the Python execution environment, triggering credential theft that collects cloud service login credentials, AI service keys, remote access private keys, and keys stored in the LiteLLM database, sending them to external servers.

During its investigation, the LiteLLM team discovered that Anthropic's coding assistant Claude Code can place GitHub login credentials in its response content under certain circumstances, and this data might be logged within LiteLLM's internal systems. It remains unclear whether the hackers directly obtained credentials this way or how they initially breached the internal systems; the investigation is ongoing. Notably, the public source code for version 1.107.0 on GitHub did not contain the malicious files, which existed only in the built release packages.

LiteLLM removed the compromised PyPI version on October 21 (uploaded around 19:39 and taken down around 21:23 Taiwan Time) and completed the removal or replacement of affected Docker Hub images by 22:20. The company has paused routine version releases. Users who installed or executed the affected versions are advised to reinstall from a clean environment, rotate potentially exposed service keys and login credentials, and temporarily revert to version 1.106.2. LiteLLM warns against running package installation commands directly in compromised environments to avoid re-triggering the malware.

출처:iThome

리서치

Nvidia Unveils DSX POD Reference Design, Claiming AI Data Center Deployment Time Cut to 14 Days

주제 · 英伟达DSX POD중대
2026-10-22 13:04 GMT+8

At GTC Berlin 2026, US chipmaker Nvidia unveiled the DSX POD, a reference design based on its latest Vera Rubin platform. The system integrates computing, networking, cabling, power, and liquid cooling into modular shipping containers, allowing partners to complete assembly and validation in factories rather than on-site at data centers.

Nvidia CEO Jensen Huang stated that this approach brings manufacturing precision and speed to AI infrastructure production, reducing deployment time from six months to as little as 14 days. By compressing the time to bring infrastructure online from half a year to under two weeks, enterprises and cloud service providers can convert large capital expenditures (Capex) into actual computing capacity and commercial revenue much faster. Global shipments are expected to begin in early 2027.

Alongside the hardware, Nvidia also released a developer preview of DSX OS, packaging its previously internal DGX Cloud operations software as open-source and modular tools. This includes network configuration management, rack power and firmware management, and cluster readiness engines, aimed at simplifying the operation of large-scale AI factories.

출처:iThome

리서치

Bank of Korea Governor: Tightening Continues Until Inflation Stabilizes; Upgrades Growth and Price Forecasts

주제 · 韩国央行货币政策중대
2026-10-22 12:16 GMT+8

The Monetary Policy Committee of the Bank of Korea (BOK) decided on October 22 to keep the base rate unchanged at 3.00% annually. This marks the first hold following consecutive rate hikes in July and August (totaling 50 basis points). Despite the pause in action, BOK Governor Shin Hyun-song delivered a strongly hawkish message in the subsequent press conference, explicitly stating that the tight monetary policy stance will persist until inflation stably converges to the 2% target level.

The core of this 'hawkish pause' lies in the BOK's significant upward revision of macroeconomic expectations. Governor Shin noted that compared to the economic outlook released in August, South Korea currently faces greater upside risks to both growth and prices. He predicted that this year's GDP growth would significantly exceed the previous forecast of 3.3%, and next year's projection of 2.9% might also be surpassed. Regarding inflation, he hinted that this year's price rise would likely exceed the previously forecast 2.9%, with similar upward pressure on the 2.8% projection for next year.

This optimistic growth and high-inflation expectation is primarily driven by exports. Shin revealed that South Korea's IT product export growth reached triple digits (262.8%) in September, pushing total monthly exports past $120 billion to a record high. Based on this, the BOK expects total customs exports for the year to reach $1 trillion, with the current account surplus exceeding the August forecast of $450 billion. Strong external demand boosts GDP while simultaneously intensifying domestic inflationary pressures.

Explaining the decision to pause rather than hike further, Shin cited the need to observe the transmission effects of the previous two 'preemptive' rate hikes and assess external uncertainties such as the Middle East situation and the US midterm elections (November 3, 2026). He mentioned that despite the Federal Reserve's rate hike in September, South Korean financial markets remained stable, and the downward stabilization of the won against the dollar helped suppress import prices, thereby enhancing purchasing power for low-income households and mitigating polarization. However, he emphasized that future timing and speed of rate hikes will depend entirely on subsequent data verification, particularly Q3 and Q4 GDP figures and October price trends.

출처:zdnet.co.kr

리서치

OpenAI CEO Sam Altman: AI Safety Coordination Requires Three-Layer Mechanism, Industry Joint Pause Needs US Antitrust Exemption

주제 · OpenAI安全协调框架중대
2026-10-05 19:00 GMT+8

In a video interview published by Vanity Fair on October 5, 2026, OpenAI CEO Sam Altman responded to host Mark Guiducci’s question about whether he would press a hypothetical 'red button' to pause AI development alongside competitors in the United States and China. His answer was not a simple yes or no; instead, he deconstructed the issue into three layers: unilateral action, industry collaboration, and international agreements. The second layer—industry-led deceleration—explicitly required an exemption from United States antitrust law, as current regulations prohibit companies from coordinating to 'slow down together.' This represents the most specific institutional barrier mentioned in his response, one less frequently cited in other AI safety discussions.

Altman described a mechanism where, at each new level of capability, sufficient progress in alignment, monitorability, and safety must be achieved before safely advancing to the next tier; if these conditions are not met, even unilateral pauses are necessary. He cited OpenAI’s recent suspension of a model training run because capabilities advanced too rapidly without corresponding improvements in alignment and monitorability. However, he acknowledged that unilateral action is insufficient because 'other open models will exist, and people can use them to do very terrible things,' which 'would require enormous sacrifices of liberty to stop.' This logic underpins his three-tier framework: unilateral pauses only control proprietary models, not the open-source ecosystem.

Altman suggested that the present moment might be a window for driving industry coordination. He noted that his stance has been 'remarkably consistent' over the past decade, but now more people believe his assessment, meaning there 'may be public support to do this.' He mentioned that OpenAI is currently in dialogue with safety teams and leaders at other laboratories and with the United States government. He envisioned that if the United States and China could agree on shared safety standards and compliance verification requirements, it would be an 'incredible step.' However, he did not specify the current stage of these dialogues or provide a timeline.

The most obvious limitation of this framework lies in verification. Altman claimed 'there have been many pauses, you just haven’t noticed,' but external observers cannot verify the frequency, triggers, or actual effects of these pauses. A more fundamental question remains: who sets the threshold for 'sufficient safety progress'? Altman stated that 'society must decide on the trade-offs here,' but did not explain how such decisions are made, who executes them, or how regulatory capture is avoided when the threshold setters are themselves stakeholders. He left this to a combination of 'governments and companies,' but the interview did not elaborate on the specific structure of this partnership.

출처:Vanity Fair

리서치

OpenAI and Google's Wiz Use AI to Find Critical PHP Flaws; US NSA Warns Millions of Sites to Patch or Go Offline

주제 · PHP重大漏洞구조적
2026-10-22 01:59 GMT+8

The U.S. National Security Agency (NSA) issued an urgent warning on October 21, 2026, regarding two critical vulnerabilities found in PHP, a server-side programming language. Because PHP powers approximately 70% of websites globally, including major content management systems like WordPress, this discovery has raised widespread concerns about internet infrastructure. The NSA advised that if a website cannot immediately apply patches, it should be temporarily pulled from the web until repairs are complete.

These vulnerabilities were discovered by OpenAI and cloud security firm Wiz (owned by Google) using artificial intelligence. OpenAI's security testing partner Calif used the GPT-5.6 Sol model to locate one flaw as part of its "Patch the Planet" initiative; the Wiz team combined its Wiz Atlas security AI agent, Gemini 4 Argon, and GPT-5.6 Sol to find multiple PHP vulnerabilities. Notably, these defects have existed since PHP version 5 was released in 2004 but were previously too complex for hackers to exploit effectively.

Wiz noted in a blog post that AI has changed the game: "LLMs can help adapt an exploit to different PHP versions, configurations and memory layouts." This means attack methods that previously required extreme expertise may now be automated via AI. Although OpenAI stated that researchers found these flaws before hackers did and there are no reports of exploits in the wild yet, public disclosure means hackers are now aware. The internet faces a race against time to patch tens of millions of websites before attacks occur.

The PHP Foundation has released patches for current versions (such as 8.5.12). Websites running older PHP versions (8.1 and earlier are no longer supported) need to upgrade. This incident highlights the double-edged sword of AI in cybersecurity: it helped defenders discover and patch potential catastrophic vulnerabilities early, but also lowered the technical barrier for attackers to exploit deep system flaws.

출처:TechCrunch

리서치

Meta Plans New Non-Voting Stock Class to Fund AI Expansion While Preserving Founder Control

주제 · Meta无投票权新股融资구조적
2026-10-22 06:45 GMT+8

US social media giant Meta confirmed during its earnings call on Wednesday that it plans to create a new class of non-voting capital stock. Chief Financial Officer Susan Li stated that these shares would be distributed to all existing shareholders via a stock dividend, aiming to provide additional funding tools for artificial intelligence investments, employee compensation, and acquisitions without diluting the voting rights of Class A and B shareholders.

This move directly addresses Meta's massive AI spending requirements. Li noted that the new shares will "expand our funding tools and give us flexibility to invest behind the unprecedented momentum we're seeing across every part of our business, driven by AI." Currently, Meta is heavily investing in the development of AI models and services, including Muse Spark and the coding agent Muse Code for developers. Although the company reported $64 billion in revenue for the third quarter, a 26% year-over-year increase, its costs and expenses also surged by 68% year-over-year, highlighting the high cost of AI construction.

By introducing non-voting stock, Meta attempts to open more equity financing channels to the public market while maintaining its "founder-led structure." This means investors can gain economic exposure to Meta without influencing corporate governance decisions. Such structures are not uncommon in the tech industry (like Alphabet), but in the current AI arms race, they offer a more flexible balance sheet management tool for large-scale capital expenditures.

Additionally, Meta disclosed that losses from its Reality Labs division (responsible for AR/VR) are expected to narrow from 2025 levels to between $18 billion and $19 billion. CEO Mark Zuckerberg emphasized that its AI assistant product Muse has been set up by over 10 million users in the US and Canada, claiming the product has entered a scaling operational phase.

출처:TechCrunch

리서치

Meta Raises 2026 AI Capex Floor to $135 Billion as Q3 Revenue Hits Record

주제 · Meta千亿AI投入策略구조적
2026-10-22 23:06 GMT+8

Social media giant Meta Platforms disclosed that it has raised the lower end of its 2026 capital expenditure (Capex) forecast to $135 billion from $130 billion, while keeping the upper bound unchanged at $145 billion. This adjustment reflects the company's accelerating investment in artificial intelligence infrastructure, custom chips, and data centers. According to a report by Proactive Investors, Meta's third-quarter revenue reached $64.6 billion, surpassing Wall Street expectations of approximately $63.4 billion, indicating that its core Facebook and Instagram advertising business continues to generate sufficient cash to support its massive AI ambitions.

Despite strong revenue, this scale of capital expenditure is reshaping Meta's financial profile. As data center construction and computing costs rise, the enormous free cash flows historically generated by social apps are coming under pressure. CEO Mark Zuckerberg has reorganized significant parts of the business around AI, attempting to compete directly with OpenAI, Google, and Anthropic. Beyond improving existing advertising systems, Meta is building standalone AI products, aiming to create new revenue streams beyond traditional advertising in the future.

Meta is far from alone. Technology giants like Microsoft, Alphabet, and Amazon have also committed hundreds of billions of dollars to AI-related infrastructure, creating enormous demand for semiconductors, electricity, and networking equipment. However, this expansion has raised concerns about whether grid infrastructure can be built quickly enough to keep pace with projected computing demand. For Meta, the central question remains whether spending approaching $145 billion a year will establish a durable competitive advantage or eventually become a drag on shareholder returns. Currently, strong revenue growth provides some buffer space.

출처:Proactiveinvestors UK

리서치

US Legal AI Firm Legora Raises $550M Series E at $8.8B Valuation, Forms Korea Team

주제 · Legora融资动态중대
2026-10-22 10:59 GMT+8

US-based legal AI company Legora announced on October 22, 2026, that it has closed a $550 million Series E funding round, bringing its valuation to $8.8 billion. This represents an approximately 57% increase from its previous valuation of $5.6 billion. The round was led by Atreides Management, with participation from new investors including Baron Capital, Capital Group, Fidelity Management & Research, and JPMorgan. Existing investors such as Accel, Benchmark, and General Catalyst also followed on.

Legora develops an AI agent platform for legal professionals, supporting tasks like legal research, document review, and drafting. The service is currently used by over 2,400 law firms and corporate legal teams across more than 80 countries, with over 140,000 active legal professional users. The company's annual recurring revenue (ARR) has grown rapidly, reaching $100 million from $1 million in 18 months, and surpassing $200 million just six months later. In the third quarter of 2026, corporate legal teams accounted for more than 40% of new customers.

The newly raised capital will be used to expand operations in the US and Asia-Pacific (APAC) regions and to enhance its 'Legora aOS' legal AI platform. The company plans to extend agent capabilities, including automated intake and classification of internal legal requests, regulatory change detection, and AI-powered legal research. Legora is also accelerating its entry into the South Korean market. Since officially launching there in July 2026, it has introduced a Korean user interface, support for Hangul Word Processor (HWP) files, and integration with local legal databases. With increasing inquiries from South Korean law firms and corporate legal departments, the company has begun establishing a dedicated team for the region.

Heather Patterson, Vice President for APAC and Japan at Legora, stated that the South Korean legal market has high receptivity to new technologies and rapid adoption speeds in practice. She noted that the company will actively invest in expanding personnel and refining products tailored to local legal needs through the formation of this dedicated team. Legora plans to grow its APAC workforce to over 100 employees by the end of 2026, with existing offices in Seoul, Sydney, Singapore, Tokyo, and Bangalore.

출처:zdnet.co.kr

리서치

OpenAI Models Land on Google Cloud, Purchases Offset GCP Commitment Credits

주제 · OpenAI与谷歌云合作중대
2026-10-22 12:10 GMT+8

US-based AI company OpenAI announced on October 21, 2026, a partnership with Google Cloud to integrate its top-tier models—GPT-6 Astra, GPT-6.1 Sol, and GPT-6 Luna—along with Codex agent workflows into Google Cloud's Model Garden platform. This move marks a significant expansion of OpenAI's cloud sales channels beyond Microsoft Azure to major global public cloud providers.

According to the announcement, the three models are available in private preview to select enterprise customers starting October 21, with general availability (GA) expected within weeks. A key financial mechanism for enterprises already holding Google Cloud commitment spend contracts is that purchases of OpenAI models via Google Cloud Marketplace can directly offset their existing GCP commitment credits. This allows companies to utilize their current cloud budgets for OpenAI models without initiating separate procurement processes.

Michael Gerstenhaber, VP of Product Management for Google Cloud Gemini, noted that customers can build applications using these models via API and connect local Codex CLI and desktop workflows directly to the Gemini platform for inference. Greg Brockman, President and Co-founder of OpenAI, stated that this partnership brings OpenAI's most powerful models into customers' existing deployment environments, offering more choices. Thomas Kurian, CEO of Google Cloud, emphasized the goal of providing diverse model options to help customers achieve optimal quality and lower costs.

Previously, OpenAI models were made available on Amazon Web Services (AWS) Bedrock and Oracle Cloud. Although OpenAI reiterated in its announcement that 'Microsoft remains OpenAI's primary cloud partner,' the collaborations with AWS, Oracle, and now Google Cloud demonstrate OpenAI's strategy to build a multi-cloud distribution network, catering to enterprise needs for avoiding vendor lock-in and flexibly utilizing existing cloud budgets.

출처:INSIDE

리서치

Walmart Launches Agent Gateway and Profiles to Enable Controlled AI Shopping

주제 · 沃尔玛AI代理购物중대
2026-10-22 03:30 GMT+8

U.S. retail giant Walmart announced on October 21, 2026, the launch of "Agent Gateway," a new infrastructure designed to help authorized AI agents shop on its website. This move marks a formal embrace of "agentic commerce" by a major retailer, standing in stark contrast to competitor Amazon's recent decision to block AI agents.

Agent Gateway is a server based on the Model Context Protocol (MCP), originally developed by Anthropic to allow AI agents to connect to external data sources and tools. Through this gateway, approved AI agents can search Walmart's online inventory and make purchases on behalf of users. Walmart stated that the technology will also support agentic shopping at its discount warehouse chain, Sam’s Club.

To address concerns about privacy and control, Walmart introduced two key accompanying measures. First, the Agent Gateway is designed to help customers shop without sharing personal information or purchase history directly with the agent. Second, "Agent Profiles," rolling out next month, will allow users to create dedicated profiles for each AI agent they use within their account. Users can view agent transactions separately and set controls, such as requiring approval for each transaction or setting spending limits, similar to managing an authorized user on a credit card.

Additionally, Walmart has established special "terms" for AI agents, defining their operational rights and liability boundaries. Notably, Walmart committed that its existing returns process applies to transactions made by AI agents, even if the agent misbehaves. This policy lowers the risk barrier for consumers trying new technology. The terms also stipulate that agents cannot access payment information and must only retrieve data necessary for the task; non-compliant agents will be blocked from accessing the site.

Walmart President and CEO John Furner stated at the Masters of Scale Summit in San Francisco that these technologies are preparing for a world where AI agents become mainstream. Previously, Walmart had partnered with Meta’s Muse agent and integrated with popular AI apps like ChatGPT and Gemini. This update further solidifies its open stance in the AI agent ecosystem.

출처:TechCrunch

리서치

UK AI audio unicorn ElevenLabs launches Realtime models to simplify voice agent development

중대
2026-10-22 11:40 GMT+8

On October 20, UK-based AI audio company ElevenLabs launched the Eleven Realtime series, including v1 Mini for natural conversation and v1 Max for complex tasks. The core innovation lies in simplifying the development of Voice Agents: previously, developers had to separately select and tune three independent components—Automatic Speech Recognition (ASR), Large Language Models (LLM), and Text-to-Speech (TTS). The Realtime series encapsulates these underlying capabilities into a unified interface, allowing developers to choose between Mini or Max based on business scenarios and directly configure roles, knowledge bases, and tool-calling logic without manual component stitching.

According to independent tests by third-party AI evaluation firm Artificial Analysis, Eleven Realtime v1 Max achieved a 98.8% accuracy rate in the Big Bench Audio speech reasoning evaluation, with an average first-audio response time of 0.57 seconds. In this specific ranking, Max placed third, behind StepAudio 3 Realtime from StepFun (99.7%) and Qwen Audio 3.0 Realtime Plus from Alibaba (99.2%). Notably, as a reference metric, Qwen Audio 3.0 Realtime Plus has a first-audio response time of 1.54 seconds, highlighting ElevenLabs' competitiveness in low-latency interaction.

In the Speech Agent Arena evaluation involving real human conversations, Max demonstrated a tool-use task completion rate of 87.1%, significantly higher than Mini's 60.3%, indicating that Max is better suited for multi-step workflows requiring frequent external API calls. However, in the Full-Duplex-Bench evaluation, Mini scored slightly higher (95.8%) than Max (94.1%) in comprehensive metrics such as pause handling and interruption response, confirming the official positioning of Mini as a solution for 'low-cost natural conversation'.

Hands-on testing by Zhidongxi showed that using voice commands via the Architect assistant to modify Agent settings indeed reduced menu navigation, but Chinese speech output carried a noticeable foreign accent, with verbose responses and occasional lag. Technically, ElevenLabs continues to use a cascaded architecture, where independent ASR, LLM, and TTS modules work together. The company acknowledges that while this architecture facilitates individual upgrades and error troubleshooting, the conversion from speech to text loses some tonal and emotional information, which requires additional modeling to compensate. Currently, both models are accessible via the ElevenAgents platform or APIs compatible with the OpenAI Realtime specification, priced by connection duration and token usage.

출처:zhidx.com

리서치

Tesla Starts Optimus Robot Production in California as Profit Falls 43%

주제 · 特斯拉Optimus量产중대
2026-10-22 04:33 GMT+8

U.S. electric vehicle maker Tesla announced on Wednesday, October 21, 2026, that it has started production of its Optimus humanoid robot at its factory in Fremont, California, utilizing a manufacturing space previously dedicated to electric vehicles. In its third-quarter letter to shareholders, Tesla stated, “The bots coming off this line are the first production-quality bots that we have manufactured.” The company noted that while hardware and manufacturing processes continue to iterate, they have proven the current process is suitable for a commercial product.

This milestone comes with significant financial pressure. Tesla CEO Elon Musk described this as “early production” during the earnings call, warning that the ramp-up will be slow and “somewhat bumpy because it is such a radically new product.” He claimed Tesla could reach 20,000 units per week by the end of 2027, but acknowledged ongoing needs to tweak design for safety, reliability, and longevity, as well as substantial work required in AI to make the robot truly useful for open-ended tasks.

Aggressive investment in AI and robotics directly impacted short-term profitability. According to TechCrunch reporting and Tesla’s disclosures, third-quarter revenue remained flat at $28.04 billion, slightly below the $28.09 billion generated in the same period last year. However, net income fell to $787 million, a 43% drop from $1.37 billion in the third quarter of 2025, driven by skyrocketing expenses for Optimus robots, AI projects, and autonomous driving. Lower regulatory credit sales also pressured profitability.

Core automotive performance was mixed: Tesla delivered 486,532 vehicles in Q3, an increase over the previous quarter but fewer than in the same period last year. Revenue from solar and energy storage declined 16% year-over-year to $2.8 billion, attributed to tariff-related customer refunds. Conversely, the services business, which includes Full Self-Driving (FSD) subscriptions, grew 33% year-over-year to $4.6 billion, setting a record for gross profit.

For investors, this event illustrates the typical financial profile of major tech companies transitioning AI from software to physical robotics: massive upfront capital expenditure and R&D costs rapidly erode traditional business profits, while revenue contributions from new products remain in very early stages. Musk also noted that the timeline for consumer deployment of unsupervised FSD (equivalent to Level 4 autonomy) has been delayed again, with expectations now set for Q2 next year for some private vehicles to join the robotaxi fleet, pushing back his April promise of a Q4 release.

출처:TechCrunch

리서치
다음 →
로그인 / 회원가입
독서리서치레이더투자 프레임워크