LeituraPesquisaRadarFramework de investimento
Entrar / Cadastrar
TudoTecnologiaProdutosInvestimentoBolsaOutros

Leitura

2026-11-1623 posts

Atualizar

a16z Author Zane Lackey: AI Is Bringing Vulnerability Discovery Back to the Pace of the 1990s

Tópico · AI漏洞发现加速Material
2026-11-16 23:02 GMT+8

In May 1998, seven hackers from L0pht testified before a U.S. Senate committee that they could take down the Internet in about 30 minutes. Lackey points out that no one in the security community at the time found this unbelievable—the state of affairs was simply like that. In 2002 alone, serious remote vulnerabilities struck foundational software supporting the Internet, including Apache, OpenSSH, OpenSSL, and BIND. Lackey believes this cadence is returning: vulnerabilities that used to appear once every few years, requiring entire security teams to drop their work for emergency response, are becoming the norm for most teams in 2026.

Lackey’s core judgment is that Opus 4.5/4.6 marks a fundamental shift in the economics of the attack side. Work that previously required scarce expertise and substantial time for vulnerability discovery and exploit development is becoming orders of magnitude faster, cheaper, and easier to repeat. The results are already appearing in key software packages that have been audited repeatedly over the past two decades—or so we thought. He illustrates the dual nature of the problem with a long-standing joke among security teams: zero-days always attract all the attention, but it is the 550-day vulnerabilities that actually cause most breaches. This joke has had a remarkably long shelf life, but now teams must worry about vulnerabilities at both ends of the time spectrum simultaneously.

Based on this judgment, Lackey proposes three structural predictions. First, Internet worms will return. Early 2000s worms exploited known vulnerabilities to self-propagate, sometimes paralyzing entire networks due to the sheer number of infected hosts scanning for new ones; this style of attack dropped significantly in the late 2000s as the Microsoft ecosystem became more secure. When AI drastically shortens vulnerability discovery and exploit development time while patch timelines lag behind, the window for worms reopens. Unlike SQLSlammer and Sasser, this round of attackers will have more commercial goals—such as stealing API keys for inference (rapidly displacing crypto mining as the preferred action for using stolen resources) and ransomware.

Second, the patching problems of the Windows XP era will recur. Unpatched Windows machines connected directly to the Internet in the early 2000s were compromised before they finished downloading updates. Lackey argues that the AI-era version of this problem extends beyond operating systems to the entire software ecosystem. Enterprises previously delayed automatic updates for two reasons: first, the risk of update-induced failures outweighed the security risk of not updating; second, the problem is far more complex than it sounds, given the staggering amount of software within enterprises, much of which is no longer actively maintained. AI unlocking offensive capabilities at machine speed will change this calculus, forcing enterprises to patch at speeds orders of magnitude faster.

Third, a new generation of startups will emerge around hardened infrastructure components. In previous similar environments, projects like OpenBSD built their identity around correctness, security auditing, and secure defaults. Lackey expects a new wave of products whose core promise is delivering highly hardened versions of critical infrastructure components for enterprise use. While previous generations of such products struggled commercially, changes in the security landscape may unlock a significantly larger market this time.

Fontes:a16z.news

Pesquisa

Mazama Energy Develops Carbon Chill CO2 Cooling Tech, Claims 20% Faster Geothermal Drilling

Visão rápida
2026-11-16 22:00 GMT+8

US-based geothermal startup Mazama Energy has exclusively shown TechCrunch its new cooling technology, called "Carbon Chill." The system injects liquid carbon dioxide down the center of the drill string, where it becomes supercritical due to surrounding rock heat and then rapidly cools as it expands through a nozzle near the drill head, absorbing heat from the bit and its electronics.

Sriram Vasantharajan, co-founder and CEO of Mazama Energy, stated that the technology increased their rate of penetration by 20% over already high baseline speeds, enabling operations in deeper, super-hot rock formations.

Drilling typically accounts for 25% to 30% of the total cost of building a geothermal power plant. Traditional mud-based cooling systems struggle with extreme temperatures, leading to electronic failures in the drill bit. Mazama aims to use this technology to prove the economic viability of super-hot rock geothermal energy (temperatures exceeding 400°C). The company, backed by Khosla Ventures, raised $135 million in September 2026.

Fontes:TechCrunch

Pesquisa

US Commerce Dept Issues New Rule Requiring Licenses for AI Chip Exports to Global Data Centers to Block Remote Compute Access

Tópico · 美国AI芯片出口管制Estrutural
2026-11-16 23:02 GMT+8

The U.S. Department of Commerce released an interim rule on November 16, 2026, significantly tightening export controls on artificial intelligence chips. The core change is that restrictions are no longer limited solely by the physical destination of the hardware. Instead, companies must now obtain licenses to export, reexport, or transfer AI chips and servers to data centers anywhere in the world if those facilities may provide computing power to companies in China or other restricted countries. This move aims to prevent sanctioned entities from circumventing existing hardware embargoes by remotely accessing infrastructure located in third-party countries.

The rule targets companies based in the U.S.'s D:5 group of embargoed countries, which includes China, Russia, Iran, North Korea, and others. Under the new regulations, any export serving a network or data center with total processing performance (TPP) exceeding 12 million requires licensing. This threshold roughly corresponds to the maximum computing power of about 750 Nvidia H100 or H200 chips. Additionally, the rule explicitly covers indirect access through go-betweens or brokers, requiring exporters and data center operators to monitor their customers' ownership structures and activities to ensure they are not controlled by anyone from the D:5 group.

Beyond global remote access restrictions, the new rule adds licensing requirements for chip exports to Malaysia, Singapore, Indonesia, Thailand, and Brazil. The Commerce Department stated these countries present a "higher-than-normal risk of diversion." However, these specific country requirements do not apply to chips destined for companies based in the U.S. Existing data center operators falling under this new rule must obtain licenses within 90 days of its effective date. Notably, the rule does not apply to commercial recommendation systems used in e-commerce or social media.

This new regulation also rescinds the Biden administration's AI Diffusion Rule from January 2025. The Trump administration had previously criticized that framework for adding unnecessary restrictions on U.S. companies and allies. The new rule emphasizes its goal to prevent the use of certain Super Intelligence (SI) items at computing power levels capable of training or running advanced SI models, thereby avoiding the use of such compute for military and intelligence-related activities that could threaten U.S. national security or foreign policy interests.

Fontes:TechCrunch

Pesquisa

Semiconductor Equipment Maker Lam Research: Using Front-End Mass-Production Technology to Break Through AI Chip Packaging Bottlenecks

Tópico · Lam ResearchMaterial
2026-11-16 18:30 GMT+8

Performance gains in AI chips have long relied on transistor miniaturization, but this path is approaching physical limits. Meanwhile, the widespread adoption of AI has caused data volumes to grow explosively, requiring larger-capacity, wider-bandwidth memory, while system power consumption and heat dissipation continue to rise. At a media briefing on November 10, 2026, semiconductor equipment maker Lam Research proposed that advanced packaging is the new partner to address these physical constraints. The strategic logic is: since relying solely on front-end miniaturization can no longer boost AI chip performance, technological breakthroughs in the packaging stage have become a complementary source of incremental gains.

Chee Ping Lee, Managing Director of Advanced Packaging Strategy Marketing at Lam Research, concretized this argument into three technical directions: shifting from horizontal layout to Z-axis stacking 3D packaging, high-aspect-ratio silicon-through-hole (TSV) etching, and co-packaged optics (CPO). The company emphasizes that its core advantage lies in directly transferring large-scale mass-production technologies accumulated in the front-end field—wet processing, thin-film deposition, and plasma etching—into packaging processes. This means packaging is no longer an independent back-end domain, but rather an extended application of front-end capabilities.

In the direction of 3D packaging, Lam made specific technical performance claims: TSV etching requires an aspect ratio of over 15:1, and its plasma etching technology achieves high-precision control by increasing cycle speed; in the chip dicing stage, plasma dicing significantly reduces particle contamination compared to traditional blade methods, while achieving more than 1.5 times the chip strength. Chee Ping Lee explicitly pointed out that improved dicing precision directly leads to better yield, and because AI chips are extremely expensive, yield improvements translate directly into economic value. This economic logic directly links packaging technological progress with the manufacturing cost structure of AI chips.

It should be noted that the above specific performance figures all come from Lam Research's own media briefing and belong to vendor-reported data, which has not yet been independently verified by third parties or subjected to peer comparative testing. Lam's strategic positioning of migrating front-end capabilities to packaging is also a self-narrative of an equipment manufacturer; actual customer adoption rates, comparisons with competitors' packaging solutions, and the performance of this technology route in mass-production environments were not provided with independent evidence in this material. However, its argumentative structure—physical limits driving packaging innovation, front-end capabilities providing differentiated entry points, and yield improvement constituting an economic lever—provides a concrete analytical framework for understanding the evolving role of equipment makers in the AI chip supply chain.

Fontes:EE Times Japan

Pesquisa

Nvidia's CUDA Moat: How AI Code Porting Undermines Software Lock-In

Material
2026-11-16 19:00 GMT+8

The most formidable competitive barrier Nvidia built over the past decade was not its chips themselves, but CUDA—a software architecture that allows developers to easily write programs and build AI models for Nvidia hardware. Once developers invest significant engineering time into CUDA, the cost of migrating to other hardware platforms is extremely high, creating what is known as a 'moat.' In an analysis published on November 16, 2026, WIRED senior writer Will Knight points out that this moat is being eroded by the very technological wave Nvidia itself has driven: AI models have become proficient enough at reading and writing code to port AI models onto competitors' chips within hours.

A concrete example of this mechanism is Wafer, a startup company. Emilio Andere, CEO of Wafer, stated that whenever a major open-weight model is released, Wafer can get it running on AMD hardware at least 80% efficiently within hours. Andere noted that AI has become so capable that his engineers rarely need to edit or review the code. This means that manual porting work, which previously took weeks or even months, is now drastically compressed by AI automation. For Nvidia, which relies on the CUDA lock-in effect, the friction costs for developers to migrate are dropping sharply.

At the same time, the rise of AI agents has made inference—running already trained models—increasingly important. Patrick Moorhead, chief analyst at MoorInStrat, pointed out that the emergence of AI agents has heightened the importance of inference, which is precisely an area where Nvidia holds a weaker advantage compared to hyperscale cloud providers and startups. Google, Amazon, and Microsoft are all developing more specialized chips to drive training and serving services for AI models, each gaining similar market share increases. Although Nvidia acquired a specialized chip startup named Grok in December 2025, it still faces structural shortcomings in inference and general-purpose CPU domains.

Nvidia's contraction in the Chinese market constitutes another layer of pressure. According to Nvidia's own filings, China's revenue share has fallen from 20% in 2024 to 8% in 2026. Washington has restricted exports of the most advanced chips to China, while Beijing pushes for the development of its domestic chip industry. Huawei has developed new types of supercomputers that utilize many more chips working in coordination. Chinese AI company Z.ai claims to have fully trained its models using Huawei chips. If Chinese open-weight models gain adoption in other countries, they could create overseas markets for Huawei chips.

However, whether these pressures constitute an existential threat remains controversial. Moorhead acknowledges that Nvidia also possesses AI tools and is not lagging behind in AI chip design. Dan Ives, partner at Yorkville Ives, stated in a memo to investors that Nvidia remains about two to three years ahead of its nearest competitors, and its strategic roadmap, driven by 42,000 employees, will not change due to the founder's departure. But chip industry analyst Dan Hutchinson likened Nvidia's current situation to Intel's 'Bob Swan moment'—in 2019, after engineering-focused CEO Brian Krzanich resigned amid an employee relationship scandal, CFO Bob Swan took over, shifting the company's focus toward sales, leading to bureaucratic bloat and declining revenue. Hutchinson believes 'Nvidia is now an Intel waiting to happen.'

Fontes:WIRED

Pesquisa

SolidWorks CEO Manish Kumar: As AI Generates More Designs, Verifying Correctness Becomes More Valuable

Tópico · SolidWorksMaterial
2026-11-16 17:44 GMT+8

In an interview with a Korean press delegation in November 2026, Dassault Systèmes SolidWorks CEO Manish Kumar offered a clear assessment of the relationship between AI and engineers: AI will make design creation extremely easy, but as the number of outputs increases, proving which result is correct becomes increasingly important. He believes this is precisely where the value of engineers emerges.

This assessment is not a vague 'AI will not replace humans.' Manish Kumar provided specific mechanisms. He used the REACT (Reasoning and Acting) framework as an example to explain how AI works in the design process: when a user requests an engineering drawing, the AI faces nearly infinite combinations of possibilities regarding view selection, dimension placement, and section arrangement. The AI does not directly output a single answer; instead, it uses a reasoning model to decide 'why dimensions are placed in this view,' executes the action, observes the result, and adjusts as needed. This process itself is iterative and autonomous.

But here lies the key turning point: AI can execute this iterative process only if a human tells it what is required. Manish Kumar used the example of designing a lawnmower robot. Suppose an engineer receives a task: reduce the robot's weight by 15% without increasing costs while maintaining its fatigue life. This requires modifying the design, running simulations, and adhering to cost constraints. He said that most engineers know how to do this, but it takes time because it involves numerous commands, clicks, and procedural steps.

AI can take over these steps. However, Manish Kumar argues that the core value of engineers has never been in 'clicking commands,' but in 'finding the correct intent'—defining goals, setting constraints on what the model should and should not do, and then evaluating the pros and cons based on the results provided by AI.

He compares this relationship to a driver and an engine: the engineer sits in the driver's seat, and AI is the engine. The engine makes the car move, but the direction and destination are the driver's responsibility.

This argument has a specific product context. This year, Dassault Systèmes is advancing SolidWorks' generative interactive AI to the 'agentic AI' level: rather than simple Q&A, humans provide specific tasks, and multiple specialized AI agents complete the work in the background according to learned processes. This system includes three 'virtual companions'—Aura, Leo, and Marie—each corresponding to different professional domains. Leo possesses 'Competencies' such as mechanical design, performance analysis, and mechanical programming; each competency is further subdivided into 'Skills,' such as conceptual design, detailed design, and drafting.

On the commercial front, Dassault Systèmes provides users who purchased SolidWorks in the past two years with 2,100 Blue Tokens per person for AI features. Each Token is priced at five cents, deducted based on executed skills and output value rather than text volume. Manish Kumar stated that within five months of commercialization, customers had already used millions of tokens, citing one customer who used the generative drawing function to increase dataset productivity by 50-60%, with a goal of reaching 90% within six months.

Fontes:바이라인네트워크

Pesquisa

Ambarella Technical Marketing Engineer Cicalese: Perception Reports Must Explicitly Retain Position, Time, Uncertainty, and Provenance

Tópico · AmbarellaMaterial
2026-11-16 14:09 GMT+8

Ambarella Technical Marketing Engineer Pietro Antonio Cicalese proposed a warehouse inspection scenario: one robot reports a fallen cardboard box in Aisle 14, and twenty seconds later, a second robot approaches from the opposite end of the aisle, reporting an obstruction near the entrance. The server now faces an awkward question—are the two robots seeing the same box? If treated as two separate obstacles, it may close more aisles than necessary; if merged into one, it might miss a second, actual obstruction. Cicalese points out that both descriptions can be accurate, yet the coordinator still cannot make a useful distinction.

He argues that the root of the problem lies in how information is handed off from perception to application. The phrase “a fallen box is blocking Aisle 14” is useful for a supervisor glancing at an alert, but for a coordinator comparing reports, it leaves substantial work unresolved: Where along the aisle is the box? How much space does it occupy? Did the second robot see it recently, or did its report spend time waiting for a connection while becoming stale? An accompanying image can help answer some of those questions, provided the recipient can interpret its geometry and capture conditions; a language model may also extract a richer description. But Cicalese emphasizes that neither approach automatically supplies a shared position nor establishes that two observations concern the same object—those relationships need to be represented and evaluated explicitly.

Therefore, the handoff requires an agreed specification between parties: what information is provided, and what behavior is expected when it is incomplete. Software engineers often call this an interface contract. Its value is practical: the receiving application should be able to distinguish a valid position from an unknown one without guessing the sender’s intent. For the warehouse, the report needs to preserve the object’s location and occupied area in a named coordinate frame, with stated units. It also needs the capture time, separate indications of uncertainty across dimensions, and a way to trace the report back to its source observation. A readable summary can sit alongside that structured information for the supervisor to quickly review.

Position conversion itself carries error. Detectors commonly locate the box within a camera image; those pixel coordinates describe where the box appears on the image plane. The coordinator needs a location on the warehouse map, and the conversion requires more than knowing where the robot is standing. A range measurement or a valid geometric constraint (such as a calibrated floor plane) is needed to locate the relevant points in three-dimensional space. The robot’s estimated position and orientation at capture time allow those points to be transformed into the map frame. The box’s extent also needs suitable geometry; the apparent width of its image alone does not establish the space it occupies on the floor.

Cicalese cites the ROS project’s REP 103 standard, noting that inconsistent units and coordinate conventions are precisely what cause integration errors. A camera’s optical frame and a robot’s body frame use different axis orientations; naming the frame and documenting the transformation prevents two components from interpreting the same numbers as different physical locations. Once both robots report in a shared map frame, the server can compare the estimated occupied regions. Overlapping locations and compatible dimensions provide evidence for a common obstacle, but they support an association hypothesis rather than proof of identity. Two similar boxes could be close together, or a worker could have moved one between observations. When evidence is insufficient to justify merging, the coordinator should retain unresolved alternatives.

Fontes:Ambarella

Pesquisa

Sierra Releases Personal Agent Protocol (Poppy) 1.0; Shopify and Meta Muse Adopt Standard

Material
2026-11-16 15:29 GMT+8

Sierra, a company developing enterprise AI agents, announced on November 13, 2026, the public release of version 1.0 of the Personal Agent Protocol (commonly known as Poppy). This open standard defines how personal AI agents securely interact with corporate websites, APIs, and enterprise-side agents on behalf of users. According to gihyo.jp, e-commerce giant Shopify has enabled automatic support for the protocol across all its stores, while Meta's AI agent Muse has made Poppy available to all its users.

The core of the Poppy protocol addresses the challenges of identity verification and permission control when AI agents access business services. Under the released 1.0 specification, companies must publish a /.well-known/poppy.json file on their domain detailing the endpoints and functions available to agents. Personal AI agents read this file to choose connection methods, including web navigation, API calls (supporting MCP and OpenAPI standards), or dialogue with the company's own agent. For account access, the protocol relies on OAuth, allowing users to grant specific read-only or write permissions and revoke them later.

The update also introduces an extension mechanism, notably the Universal Commerce Protocol (UCP) extension co-developed with Shopify. This extension combines Poppy's connection discovery with UCP's handling of product search, cart management, and payments, enabling AI agents to shop as guests without logging in or to use customer accounts with explicit user authorization. Additionally, Sierra provided an implementation kit for enterprises, containing skills documentation to guide coding agents through system adaptation and validation tools.

Beyond Shopify and Meta, financial services firm Mercury, Turkish Airlines, and Synchrony have begun adopting Poppy. For instance, Muse users can now query Mercury account balances or categorize expenses via the protocol, though transfers require approval based on bank rules. Turkish Airlines allows users to check flight status and mileage upgrade options through agents. Currently, over 100 companies, including OpenAI, Stripe, Visa, Walmart, and Salesforce, are listed as partners in the standard's development.

Fontes:gihyo.jp

Pesquisa

AI Infrastructure Startup Mount Thor Secures $45 Million to Build Mac Data Centers for AI Agents

Tópico · Mount ThorMaterial
2026-11-16 09:05 GMT+8

AI infrastructure startup Mount Thor has announced it has secured $45 million in equity funding to build data centers running racks of Apple computers, specifically designed to support the development and deployment of artificial intelligence agents. Founded by former OpenAI employee Peter Voell, the company is headquartered in San Francisco, United States.

As AI agents increasingly need to navigate desktop applications, manage files, and perform tasks dependent on specific operating system environments, traditional Linux cloud servers often fall short. Mount Thor's service provides remote access to full macOS environments, sparing customers the expense and operational complexity of maintaining extensive Mac fleets themselves.

The financing spans Seed and Series A rounds. CRV and Conviction co-led the Series A, while Chad Byers and Offline Ventures co-led the Seed. CSC Leasing, Enea Capital, Long Journey Ventures, and Upper90 also participated. Additionally, the company secured non-dilutive capital to support hardware purchases.

Mount Thor stated that its first data center was deployed within six months. The new capital will be used to expand computing capacity, hire staff, and establish a laboratory to test inference capabilities of open-source AI models optimized for Apple processors. The company argues that Apple's long-standing investment in accessibility and human-computer interaction makes macOS particularly suitable for AI agents requiring human supervision.

Fontes:TechRadar

Pesquisa

TE Connectivity Vice President of Product Management Brian Payne: Why AI Infrastructure Still Requires Dual Investment in Copper and Optics

Material
2026-11-16 15:00 GMT+8

The AI infrastructure industry has long been driven by a linear narrative: as data rates increase, copper interconnects will eventually phase out, giving way to full optical interconnects. At the 2026 OCP Global Summit, Brian Payne, Vice President of Product Management for the Digital Data Networks (DDN) business at TE Connectivity, offered a different perspective on this narrative. In an interview with ServeTheHome, he stated clearly that copper remains the state of the art, and the company will not halt its investment in copper simply because of advancements in optics.

Payne’s logic begins with physical constraints. He noted that with every chip generation update, bandwidth demands grow, and the challenge lies in extracting that data from compute trays and delivering it into the network. The 448G per-channel SerDes rate introduces severe thermal management and signal integrity issues, prompting the industry to seek solutions that balance signal quality, power consumption, and package size. At the October 2026 OCP Summit, TE Connectivity announced the USFP standard in partnership with NVIDIA and Cisco. While similar in form factor to current QSFP modules, USFP uses a unique compression mechanism to separate signal integrity handling from power delivery, significantly enhancing signal quality. Payne emphasized that the design also includes interfaces for cold plate liquid cooling to manage the rising power consumption of optical modules.

Inside the rack, Payne identifies copper backplanes and midplanes as key enablers for AI scale-up. The midplane design allows compute trays to connect to GPUs via high-speed links after separation, while copper backplanes handle switching connections on the rear side of the rack. He described TE Connectivity as one of the few companies in the industry capable of engineering such highly complex products and achieving mass production. This end-to-end coverage—from near-chip packaging and front-end I/O to backplanes and optical modules—constitutes what he calls a 'whip to chip' systems thinking approach.

Regarding optics, Payne acknowledges their necessity but rejects the notion that they will replace copper. He mentioned that TE acquired the optics company RAM Photonics earlier this year and is making massive investments in both co-packaged optics (CPO) and co-packaged copper. Addressing the long-standing discussion around CPO deployment, he explained that while people argued for CPO ten years ago, physical constraints—specifically, the fact that copper routing distances shorten as channel counts increase—are making the tipping point increasingly real. However, he stressed that this does not mean copper investment will cease, as copper remains the optimal method for delivering the results customers expect.

On the power delivery side, Payne presented a specific system-level innovation: TE’s rack design features an independent sidecar power rack that supplies adjacent IT racks via horizontal liquid-cooled busbars. This horizontal busbar is an industry first, designed to solve a practical constraint: AI rack power density has reached a point where all power equipment cannot fit within a single rack. The horizontal busbar allows power transmission across racks without increasing current or heat dissipation. He further pointed out that the next frontier is 800V high-voltage direct current (HVDC) distribution, which is necessary to achieve megawatt-class power per rack. While TE’s liquid-cooled busbars are already deployed in the industry, 800V rack-to-rack cables remain in the prototype stage.

Fontes:ServeTheHome

Pesquisa

NTT West and Iris Ohyama Verify Technical Feasibility of AI-Driven Multi-Robot Facility Management

Tópico · NTT西日本多机器人平台Visão rápida
Verificado 2026-11-16 15:35 GMT+8

NTT West Corporation and Iris Ohyama Inc., a Japanese manufacturer of household goods and appliances, announced on November 16, 2026, the results of the second phase of their joint demonstration of a "Multi-Robotics Platform." The test aimed to verify the feasibility of using artificial intelligence to unify the scheduling of robots from different manufacturers for facility management.

The core of this demonstration lies in a three-layer AI architecture: first, a Vision-Language Model (VLM) analyzes footage from building security cameras to understand the site situation; next, a Large Language Model (LLM) issues instructions to specific equipment or robots based on this context; finally, a Vision-Language-Action (VLA) model enables the robots to autonomously judge and execute actions according to the instructions and surrounding environment.

In specific scenarios, the system successfully closed the loop from capturing images via cameras to directing Iris Ohyama's cleaning robot "JILBY" to perform tasks. Additionally, the companies verified information sharing and task division between JILBY and ugo's humanoid robot "ugo Nova." While the technical path has been proven, both companies stated that commercialization and service availability timelines remain undetermined and are currently under evaluation.

Fontes:プレスリリース・ニュースリリース配信シェアNo.1|PR TIMES

Pesquisa

Google's Project Zero Announces Shorter Vulnerability Disclosure Timelines Due to AI-Accelerated Exploits

Material
2026-11-16 14:19 GMT+8

Google's vulnerability research team, Project Zero, announced on November 16, 2026, that it will adjust its disclosure policy to gradually shorten the time between reporting a vulnerability to vendors and making details public. The new policy reduces the current 90-day window to 60 days starting February 2, 2027, and further to 30 days starting January 3, 2028. After these deadlines, technical details will be published regardless of whether vendors have completed patches.

The primary reason for this adjustment is that large language models (LLMs) and other AI tools have changed the speed of vulnerability offense and defense. Project Zero believes AI allows attackers to discover and exploit vulnerabilities faster and increases the likelihood of independent discovery. Furthermore, after vendors release security updates, attackers can use AI to analyze code differences before and after the patch, inferring the vulnerability's location and cause to develop exploits. This means the previous 30-day buffer period may no longer prevent attackers from understanding the vulnerability even if details aren't officially public.

Since 2021, Project Zero used a '90+30' policy: 90 days for vendors to patch, followed by a 30-day delay after the update release before publishing details. The new policy eliminates this buffer and most special provisions, calculating a fixed disclosure deadline from the report date. The team noted that AI has increased the volume of reported vulnerabilities, with some targets now yielding weekly discoveries instead of annual ones, making case-by-case negotiation unsustainable.

The new policy does not adjust disclosure times based on severity, arguing that while severity affects harm, it doesn't change the probability of an attacker finding the flaw. Vendors can still prioritize patches by risk, but disclosure deadlines will follow consistent rules. To address vendor concerns, the policy will be implemented in phases, with exceptions retained for cases like active exploitation or vendor requests for early disclosure.

Fontes:iThome

Pesquisa
Próxima →
Entrar / Cadastrar
LeituraPesquisaRadarFramework de investimento

For security teams dealing with this situation, Lackey offers three practical recommendations. First, test yourselves aggressively—not annual compliance penetration tests, but continuous, scope-free full red team testing using the latest tools attackers are currently succeeding with. He mentions tools like Hinoki Security, which give attackers access to specialized post-trained security models that bypass model refusal mechanisms. Second, the meaning of defense in depth is shifting: it is no longer just about layering enough controls so that if one or two fail, others still work; in the new environment, even if all layers function correctly, zero-days, new techniques, and sheer speed may bypass them. Therefore, teams need to deploy overlapping controls to increase detection probability and optimize detection and response capabilities to an AI-native speed level. Third, rebuild the security stack for the AI operating environment, similar to the migration from physical devices to cloud-native stacks—endpoint security is entering the agentic malware era, identity management needs to control agents rather than just authenticate humans, and open-source supply chains face exploding attack volumes due to agentic programming.

It should be noted that this article was published on the a16z.news platform, and Neo (endpoint AI security), Keycard (agent identity management), and Socket (open-source supply chain protection) mentioned in the text are all portfolio companies of a16z, which lends a commercial interest bias to the prediction that 'new startup categories will emerge.' Additionally, the claim that 'Opus 4.5/4.6 marks a fundamental turning point' is based on attributed observations from conversations with CISOs, rather than independent quantitative measurement. The prediction of worm resurgence depends on whether attackers can assemble exploit chains before the patch window closes—Lackey admits patch timelines have not caught up, but whether this condition is sufficient to support large-scale worm propagation remains speculative rather than an observed result.

AMD's market share rose from 2% in 2022 to between 7% and 10% in 2026, while Nvidia still accounts for 80% of sales in the AI accelerator market. Nvidia's revenue grew from $26.91 billion in January 2022 to $215.94 billion in January 2026. These figures indicate that Nvidia's lead remains massive, but the loosening of the CUDA moat, the shift in inference workloads, and the shrinkage of the Chinese market are structural forces that were already operating before Jensen Huang's personal controversy.

These figures come from vendor self-reports and a single customer case, lacking publicly reproducible benchmarks. However, Manish Kumar's core argument does not depend on these numbers: when generation costs approach zero, the scarce resource shifts from 'being able to create designs' to 'being able to judge which design is correct.' This logic applies to any engineering field using AI generation tools, regardless of the specific product.

The boundary that must be preserved is: Manish Kumar is the CEO of the SolidWorks product, and his discourse serves as both a technical viewpoint and a commercial positioning. His premise that 'AI makes design easier' assumes that their own tools can indeed achieve this—there is currently no independent verification indicating that the reliability of generative AI under complex engineering constraints has reached the level of fully replacing human execution. His argument describes a direction, not a completed fact.

Time also needs to be distinguished. Capture time tells the receiver when the scene was observed; processing completion and message arrival describe later events in the pipeline. They remain useful for diagnosing delays, but using arrival time as the observation time can make an old scene appear current. Comparing capture times across robots requires clocks with a known relationship. If clock error is large relative to the event being analyzed, apparent timing differences may be misleading. The contract should make the time reference clear and define when an observation has become too old for a particular decision—an obstacle report may remain useful for scheduling after it has become unsuitable for deciding whether a narrow route is currently clear.

Uncertainty also needs more than a single score. The system may be confident that an object is a carton while having a poor estimate of its distance. A precise depth reading may accompany an uncertain object label. Glare may affect part of an image without invalidating every measurement from it. Combining all of these conditions into one confidence value leaves the receiving application unable to respond to their different consequences. For inventory, this distinction is especially tangible: a shelf label can establish which SKU belongs at a location, but it cannot reveal cartons hidden behind an obstruction. The count should cover what the robot actually observed; unobserved portions remain unknown, potentially requiring another observation or manual count to complete the task.

Provenance makes such questions possible to investigate. The report can identify the robot, sensor, and source observation, along with the processing version used to produce it. Where access and retention rules permit, a reference can lead to an image crop or recorded frame. That evidence need not travel with every message, but its availability and eventual expiry should be understood by the recipient.

As more sensors join, more measurements need reconciling. A camera may suggest the object’s type, while radar contributes range and radial velocity. Those readings can be combined only with appropriate spatial calibration, temporal alignment, and an assessment of whether they concern the same target. A radar return with no reliable camera match can remain an object of unknown type; inventing a label would conceal the disagreement. Fusion can take place at different stages, depending on what the downstream task needs and what each processor can handle. Cicalese cites Jason Huang’s account of Ambarella’s radar demonstration: a CV3 system-on-chip in the front radar unit processes data from five Oculii radar heads, then sends a combined point cloud to a zone controller for sensor fusion. A point cloud represents individual detections by their positions in three-dimensional space. That design keeps multi-radar processing together and gives the next stage a spatial representation. Other fusion tasks may need a different boundary, depending on which measurements the receiving stage requires.

Finally, Cicalese proposes that the interface contract should be tested by replaying reports from known scenes: starting with two observations of one box, then using two nearby boxes with similar appearances; introducing an old report after a newer one; removing the range estimate, changing the coordinate frame, and simulating a clock offset. Expected results should include how the coordinator handles uncertainty in each case. A useful test checks the downstream decision: if two reports cannot be associated confidently, the coordinator should preserve that ambiguity and follow the specified scheduling behavior, such as requesting another observation. A missing distance must not become zero, and a region hidden from the camera must not become clear space. Passing a message through a parser establishes much less than this.

When two robots report from opposite ends of Aisle 14, the server should have enough evidence to make a justified association—or enough information to recognize that another observation is needed. The coordinator’s assignment remains a proposal evaluated by the robot application runtime; motion remains subject to the machine’s separately validated protective control constraints. A well-designed report improves coordination without taking over those responsibilities.

Payne’s arguments carry a clear commercial context: TE Connectivity is a supplier of both copper and optical interconnect components, so its 'dual investment' strategy naturally aligns with its broad product portfolio. His assessment—that copper remains the state of the art and optics serve as a necessary complement rather than a replacement—is highly consistent with TE’s business structure. This makes his views valuable primary material for understanding the technical roadmap judgments of major component suppliers, though they should not be taken as independent technical evaluation conclusions.

The practical value for readers lies in recognizing that industry discussions on AI interconnects often focus narrowly on chip generations or single technology breakthroughs. Payne’s explanation offers a systems-level perspective: 448G front-end connectors, copper midplanes, serviceable CPO, and 800V liquid-cooled busbars are not isolated product decisions but mutually constraining elements of a whole. Scaling AI racks is not merely a bottleneck issue for a single component; it is a joint optimization problem involving bandwidth, power, and thermal dissipation within physical space constraints.