LesenForschungRadarAnlageframework
Anmelden / Registrieren
AlleTechnikProdukteInvestierenBörseSonstiges

Lesen

2026-11-1298 Beiträge

Aktualisieren

Security Firm Claroc Finds 95% of Grok-Built Apps Publicly Accessible by Default, Exposing Patient Data and Internal Sales Pipelines

Thema · Grok应用默认公开泄露Wesentlich
2026-11-11 20:53 GMT+8

Security firm Claroc revealed in a recent report that 94.7% of web applications built using SpaceXSI's (formerly xAI) AI assistant Grok were publicly accessible without any login credentials. This finding highlights significant flaws in the default security settings of AI-assisted development platforms.

According to Claroc's examination of 37,571 online apps, 35,578 were in an "Anyone with the link" state. The root cause lies in the publish dialog within Grok's "Build" mode: on the paid account used for testing, the "Anyone with the link" option was preselected. Furthermore, each app receives its own subdomain, which gets indexed in commercial passive DNS databases after the first visit, allowing attackers to easily discover these apps through pattern searches.

The leaked data included not only personal information from ordinary users but also internal confidential documents from SpaceXSI and X. Claroc discovered 24 accessible apps created by SpaceXSI and X employees, including onboarding plans for new enterprise sales team members, a sales territory map containing 594 strategic accounts and pipeline data, and ad proposals targeting U.S. federal agencies. Some documents were explicitly marked as "confidential."

Sensitive personal health data was also exposed. Confirmed cases included patient lab values in ICU briefings, a list of over 3,500 employees at a hospital group, and fire department shift schedules. In the source code of a Grok-generated timesheet app, researchers even found 18 accounts with plaintext passwords. Claroc estimates that between 650 and 800 apps contained private data, affecting 30 organizations, six of which are Fortune 500 companies.

In response to the disclosure, SpaceXSI stated that apps are private until shared by their creator, that publishing takes multiple steps, and that employee accounts default to "Only you." However, Claroc countered that extra clicks do not help if "Anyone with the link" remains preselected at the final step. Currently, SpaceXSI has taken its 24 employee apps offline but has not removed the thousands of open apps belonging to other users and organizations.

This is at least the third privacy incident involving Grok for SpaceXSI. Previously, Forbes found more than 370,000 shared Grok conversations appearing in Google Search results, and the Grok Build terminal tool uploaded user files to servers without asking. As AI-generated code becomes widespread, such large-scale data exposure risks due to improper default configurations are becoming a focal point for the industry.

Quellen:The Decoder

Forschung

Unit 42: Indicators for AI Autonomous Penetration Tools Fail Quickly; Behavioral Features Offer More Persistent Detection Value

Thema · ARTEX AI渗透框架Wesentlich
2026-11-12 07:00 GMT+8

In a report released on November 11, Palo Alto Networks threat research team Unit 42 presented a judgment with direct operational significance for defenders: when facing AI autonomous penetration tools, indicator-based detection is failing rapidly. The ARTEX framework is frequently forked and modified, and attackers rotate infrastructure at an extremely fast pace, leaving almost no defensive window before traditional indicators such as IP addresses and user agent strings lose their value. However, based on observations from multiple intrusion investigations and managed threat hunting, Unit 42 points out that there is a class of signals more durable than indicators—namely, behavior itself.

This judgment is grounded in a structural fact: autonomous penetration frameworks rely on the foundational reasoning patterns of large language models to drive decision-making. Regardless of how code is forked or infrastructure rotated, an agent that waits for model inference to complete before initiating the next round of requests will inevitably exhibit specific network behaviors. Typical patterns observed by Unit 42 include: bursts of one to fourteen requests sent at approximately one-second intervals, followed by pauses of twenty-five to thirty-one seconds, consistent with the time an agent spends waiting for a model inference call to complete; lateral movement completed within about one minute after obtaining credentials, with verification of over two hundred client secrets within three minutes; immediate extraction of routing targets from downloaded OpenAPI specifications or source code maps; and failed requests being adjusted and retried successfully within seconds. While these characteristics appearing individually might merely indicate tool configuration anomalies or manual testing, their combination points to autonomous agents.

Based on this, Unit 42’s defensive recommendation is to build detection around operator behavior rather than solely tracking indicators. Specifically, defenders should look for canary usernames (such as placeholder values like evil[.]com or test@example[.]com) in identity provider logs, monitor the first use of cloud keys from new IP addresses, identify write-then-delete test patterns on storage buckets, and focus on early probing stages preceding later intrusions. The report simultaneously acknowledges a limitation: in at least one observed case, the attacker switched from scripted clients to a desktop browser for the most sensitive steps, suggesting human intervention or browser automation may break pure behavioral patterns. Therefore, behavioral detection is necessary but not sufficient, and must still be combined with authentication hardening, attack surface reduction, and default-deny strategies.

Quellen:Unit 42

Forschung

Gaurav Chawla, Vice President of Infrastructure Solutions at Dell Technologies: 'Data Gravity' Causes Enterprise AI Pilots to Succeed but Fail at Scale

Thema · 企业AI数据引力Wesentlich
2026-11-12 10:53 GMT+8

A 2025 report from the Massachusetts Institute of Technology (MIT) indicates that up to 95% of enterprise generative AI pilot projects have little to no quantifiable impact on revenue. The report suggests that model quality is not the primary issue facing enterprises; rather, the gap lies in how AI is integrated into operational workflows. In other words, success in a pilot does not prove that the same system will run smoothly across the entire organization.

Gaurav Chawla, Vice President of Infrastructure Solutions in the Office of the CTO at Dell Technologies, recently proposed a mechanism to explain this gap in a column article: 'data gravity.' His core argument is that most enterprises, after introducing new AI platforms, mistakenly believe that simply aggregating data in one place will allow the intelligence layer above it to function naturally. This may seem reasonable during proof-of-concept stages, but it is almost impossible to achieve at an enterprise-wide scale.

The meaning of data gravity is that the larger the dataset, the higher the cost for the enterprise to move it. Specifically, regulations require certain data to be stored only in specific locations, data sovereignty rules restrict data storage to designated regions, and governance architectures built around specific datasets over the past decades give business units little motivation to migrate data to new platforms, sometimes even lacking legal or compliant pathways to do so. More critically, when attempts are made to separate data from the applications built around it, these applications may suffer severe and costly failures.

Chawla points out that the impact of data gravity is often hidden during the proof-of-concept and pilot stages, as these projects operate only on small, pre-cleaned subsets of data. The true costs borne by enterprises emerge only after AI projects are applied to the vast majority of messy, real-world environmental data. Transmission fees are just one part; latency, bandwidth, security controls, and operational overhead generated after data is processed by AI further drive up overall costs.

The most dangerous aspect of this 'hidden tax' is its timing. Chawla emphasizes that costs often surface 12 to 24 months after procurement is completed and teams are assembled, by which time the initial results brought by AI have already been reported upward, procurement contracts are signed, team personnel are hired, and the company's development direction has been publicly committed externally. At this point, revisiting and modifying the architecture costs far more than designing for data gravity from the beginning.

Chawla believes that the value limitation of enterprise AI is half in computing power and half in data. Token usage and GPU efficiency can tell teams the cost of running models, but they cannot show whether the models possess the context needed to make good decisions. AI models never need raw data poured directly in front of them; what they need is context—the ability to find the correct records, cross-reference with policies, adhere to access rules, and operate based on the latest information, rather than relying solely on data snapshots taken at project initiation.

Quellen:TechOrange 科技報橘

Forschung

Alibaba Cloud Engineer Zheng Tian: Agent Architecture Is Converging, with the Bottleneck Shifting from Model Capability to Human Attention

Thema · 智能体架构收敛Strukturell
Verifiziert 2026-11-12 12:25 GMT+8

Zheng Tian, an engineer at Alibaba Cloud, proposes a judgment in the "AI Agent Handbook": after DeepSeek released SuperAgent, the prevailing view in the industry was that Agent design had largely converged, and he shares this view. More notably, he points out that his own team arrived at almost the exact same architecture while independently developing Collaboration Space—using the same task objects and the same call-chain organization methods. He argues this is neither coincidence nor mutual borrowing, but rather that constraints have narrowed the solution space.

The causal chain Zheng Tian presents is as follows: models can already complete individual tasks, so the bottleneck has shifted to human attention. The DSH team observed that one person can usually follow only 3 to 5 active conversations at once; therefore, the ceiling for a single task depends on the model, while the ceiling for multiple tasks depends on humans. This observation directly derives three design requirements: a need for a unified external interface (so users feel they are always collaborating with just one Agent), isolated execution units (as tasks grow longer and more numerous), and task objects capable of persisting across sessions. Once a problem is clearly defined, answers tend to converge, much like how databases ultimately moved toward write-ahead logging and operating systems toward processes and schedulers.

At the engineering level, Zheng Tian summarizes this convergence into a core principle: the model is the system's only nondeterministic component, and engineering's job is to wrap it in a deterministic shell. He found five specific manifestations of this in the DeepSeek prototype code. First, the model proposes while the program decides—the Router model saying 'accept work' is essentially just submitting a request; the deterministic core checks permissions, rules, and state before actually writing, meaning that even if the model suffers from prompt injection attacks, the worst-case scenario is merely a rejected request. Second, log first, then execute—external actions and state changes are recorded in the same transaction, allowing unfinished actions to be resumed after a crash restart, which is the Outbox pattern in distributed systems. Third, external text is always just data—content in group chats and comments sits at the lowest priority; when the Router hands off a task, its note is limited to 30 words and must not include an implementation plan, forcing the Worker to read the original messages itself. Fourth, proactivity must support staged rollout—new rules default to a dry run where the model records 'I would accept this work if execution were enabled' without actually starting; execution is only turned on after the team reviews and confirms accuracy. Fifth, people have the highest priority—one concurrency slot is always reserved for humans; even if all Worker slots are busy, human messages can still wake existing work.

Zheng Tian also explicitly points out the limitations of the current solution. DSH's load testing used simulated models, with event rates reaching only 40 per minute, so results cannot be extrapolated to arbitrary scales. He identifies at least four bottlenecks in centralization: context capacity (everything passes through it, so the window will eventually run out), decision quality (judging too many things in one round invites errors with broad impact), cost (large contexts must be reread every time; SuperAgent's cached Token read volume is 3.8 times the baseline), and blast radius (all events and permissions converge at one point, so prompt injection there affects the entire system). DSH itself admits that finer partitioning increases throughput but reduces continuity, a trade-off that will persist until the global memory problem is solved.

Quellen:Alibaba Cloud Community

Forschung

US DARPA-funded project: Caltech's Astra robot performs first autonomous in-space structural assembly

Thema · Astra太空组装机器人Wesentlich
2026-11-12 00:30 GMT+8

The Nom4d project, funded by the U.S. Defense Advanced Research Projects Agency (DARPA), has achieved a key milestone: the Astra robotic system, developed by the California Institute of Technology (Caltech), successfully performed the first autonomous structural assembly in space. This breakthrough aims to overcome the physical limitations currently facing the deployment of large structures into orbit—traditional methods require structures to be folded or disassembled on the ground, launched, and then manually reassembled.

Launched into space in March 2026, the Astra robot was tasked with autonomously constructing a 1.4-meter diameter reflector antenna using carbon fiber tubes and aluminum connectors, and covering it with a reflective film. According to DARPA, this assembly method resembles a high-tech version of children's building blocks, but the automation process proved highly challenging. Due to time constraints, the team only completed the first section of the antenna. As the mission is set to end in December 2026, the antenna may not be fully finalized, but the demonstration has sufficiently proven the feasibility of the concept of autonomous in-space assembly.

Sergio Pellegrino, the project's lead scientist, Professor of Aerospace at Caltech, and Senior Research Scientist at NASA's Jet Propulsion Laboratory (JPL), stated that this success lays the foundation for building giant mirrors for more powerful space telescopes or orbital solar power stations in the future. Andrew Detor, DARPA's Nom4d program manager, noted that the long-term goal is to enable the efficient, scalable construction of very large structures hundreds of meters or even kilometers in size on orbit, which is currently impossible to achieve through traditional launch methods.

Quellen:t3n.de

Forschung

ElevenLabs and NVIDIA Expand Partnership: Adding 15MW Compute and Deploying Sovereign Voice Agent for Poland

Thema · ElevenLabsWesentlich
2026-11-11 20:00 GMT+8

On November 11, 2026, AI audio company ElevenLabs and chip giant NVIDIA announced an expansion of their strategic partnership. The collaboration covers model integration, infrastructure scaling, and sovereign deployments, aiming to help enterprises and governments run voice agents at scale.

In terms of models, NVIDIA's Nemotron 3.5 Super VL is now available on ElevenLabs' ElevenAgents platform. This model is designed for agentic workloads requiring reasoning, tool use, and execution of complex multi-step tasks. According to ElevenLabs, in benchmarks for realistic customer support tasks, Nemotron 3.5 Super VL achieved leading end-to-end task completion accuracy at 70% lower cost compared to other leading open models.

Regarding infrastructure, to meet growing demand from its largest customers, ElevenLabs plans to add 15 megawatts of next-generation NVIDIA Vera Rubin capacity, on top of its existing NVIDIA Blackwell Ultra systems. This expansion supports large-scale deployments with strict security and data-residency requirements.

Additionally, the partners are advancing sovereign deployment cases. ElevenLabs is using its voice AI models and NVIDIA infrastructure to power a voice agent for Poland's national e-Health platform (Centralna e-Rejestracja). The agent runs on-premises using NVIDIA Confidential Computing, ensuring citizen data and audio remain protected within government infrastructure, while local deployment reduces latency for real-time public services.

Quellen:ElevenLabs

Forschung

Git Patches High-Severity Vulnerability CVE-2026-111760 Allowing Code Execution via Malicious Repositories

Wesentlich
2026-11-12 08:00 GMT+8

The Git project released patches on November 10, 2026, fixing six security vulnerabilities. The most critical is CVE-2026-111760, with a CVSS 3.1 risk score of 8.0 (High). This vulnerability stems from an integer truncation issue when Git processes index data. Attackers can exploit specially crafted long paths to replace a submodule's .git file with a symbolic link pointing to a malicious configuration or hooks directory, causing Git to execute attacker-controlled code during subsequent operations.

According to GitHub, this vulnerability is primarily triggered when users recursively clone repositories and their submodules using the --recurse-submodules option. The other five vulnerabilities cover path traversal, memory corruption, out-of-bounds reads, and unexpected file access. For instance, CVE-2026-105437 can cause heap memory corruption during large merge operations, affecting applications and code hosting services that process merges of untrusted code.

Git has released version 2.56.1 and patched versions for several older maintenance branches. GitHub.com and GitHub Enterprise Cloud have completed the fixes. Updates are also available for GitHub Enterprise Server versions 3.19.15, 3.20.11, 3.21.9, and 3.22.4. GitHub announced that updates for GitHub Desktop and the GitHub Copilot app were still being coordinated at the time of the announcement, while GitHub Codespaces and GitHub Actions will be updated soon.

Enterprises should check if applications embedding Git have received patch updates. The Git project advises users to avoid recursively cloning untrusted repositories until they have applied the updates.

Quellen:iThome

Forschung

Microsoft Reports Attackers Using Agentic AI to Scan GitHub for Leaked Credentials, Exfiltrating Data from Azure Key Vault in 5 Minutes

Wesentlich
2026-11-12 08:00 GMT+8

Microsoft Threat Intelligence Center (MSTIC) reported observing attackers using agentic AI tools to mass-scan public GitHub repositories for leaked application credentials and other sensitive data. Upon finding valid Service Principal credentials, attackers use them to log into corporate Microsoft Entra ID tenants and access Azure resources. In one case, an AI agent moved from searching for secrets to exfiltrating data from Azure Key Vault in just 5 minutes.

Unlike traditional automation scripts, these AI agents can dynamically adjust their next steps based on operational results. For example, if an API call fails, the agent attempts alternative methods. This adaptability makes such attacks harder to detect with fixed signatures. Microsoft noted that this activity has been observed since August, with multiple successful intrusions confirmed in late October.

This finding underscores the critical importance of managing application credentials and cloud workload identities. Researchers warn that any publicly exposed credential should be considered compromised, even if the associated code has been deleted from GitHub; the credential itself remains valid. Enterprises must revoke or rotate these credentials and audit their past usage. Additionally, organizations should monitor for anomalous cloud login behaviors, such as repeated login attempts in short intervals, rapid access to multiple services by the same application, and quick retries after API failures, analyzing these alongside authentication logs rather than relying solely on IP blocklists.

Quellen:iThome

Forschung

AI Cloud Provider CoreWeave Upgrades to CNCF Platinum Membership to Deepen Open Source Infrastructure Involvement

Kurzfassung
2026-11-12 02:11 GMT+8

AI cloud computing company CoreWeave has upgraded its membership status with the Cloud Native Computing Foundation (CNCF) to Platinum, the foundation's highest corporate membership tier. This move signals the company's desire to more deeply participate in shaping how Kubernetes handles the large AI workloads its customers run.

Yoshi Tamura, Principal Product Manager at CoreWeave, stated that much of the company's explosive growth traces back to the open nature of the container orchestration platform and its surrounding ecosystem. "Without [Kubernetes], I don't think we would've been able to get this kind of explosive growth," Tamura said. He noted that the upgrade indicates they "really mean it" in wanting to give back to the community and be more involved.

As AI agents reshape enterprise infrastructure, cloud providers are scaling Kubernetes inference to serve frontier AI models. CoreWeave describes itself as a cloud built from the ground up for AI rather than retrofitted for it. By working closely with open-source inference engines like vLLM and SGLang, as well as model labs, CoreWeave's team aims to solve the difficulties of running newly released models in production, ensuring patches and custom kernels are available on day-zero.

Quellen:SiliconANGLE

Forschung

Nvidia Board Preparing to Oust CEO Jensen Huang, Stock Falls Over 8%

Strukturell
2026-11-12 07:12 GMT+8

Nvidia’s board of directors is preparing to oust CEO and founder Jensen Huang as soon as today, November 11, 2026, according to Semafor, which confirmed reports by The Wall Street Journal. This move follows media disclosures detailing a second extramarital affair involving Huang, which has triggered an independent investigation into his relationship with a suspected Chinese spy.

The board has briefed the White House on its plans, per a senior administration official cited in the report. Following the news, shares of the $4 trillion chip giant fell more than 8% in after-hours trading, hitting a 52-week low.

On November 8, Nvidia’s board announced an independent investigation led by Wachtell Lipton to examine whether Huang divulged corporate secrets or endangered US national security. The inquiry was prompted by reports, including from The New York Times, that Huang had a close relationship with Alice Huang, an executive at Megaspeed who was reportedly recruited by China’s Ministry of State Security. While the board initially stated there was no reason to believe national security or company secrets were compromised, these new reports mark an escalation from investigation to active preparation for leadership removal.

Quellen:semafor.com

Forschung

NYT Reports Nvidia CEO Jensen Huang Allegedly Had Affair With Direct Report; Board Preparing to Fire Him

Thema · 英伟达黄仁勋丑闻Wesentlich
2026-11-12 06:31 GMT+8

According to a New York Times report, Nvidia CEO Jensen Huang is accused of having an extramarital affair with a direct report, escalating the scandal surrounding him. The newspaper, citing five anonymous sources within Nvidia, stated that Huang had a relationship with Kimberly Powell, the company's vice president of healthcare.

The report says the two were "captured on video being affectionate" in March at a hotel in San Jose, California, where Nvidia's GTC conference takes place. Sources claim someone filmed the moment on a smartphone, but Huang's security team "sequestered" them and "made the person an offer to sign a nondisclosure agreement."

Huang was already embroiled in a scandal after admitting to an affair with Alice Huang, an executive at Singaporean firm Megaspeed whom U.S. intelligence believes was spying for China. Nvidia's board has launched an independent investigation into that matter.

The Times noted that personal relationships are mentioned as potentially causing conflicts of interest in Nvidia's code of conduct. Powell joined Nvidia in 2008 and has been steadily promoted. Like Huang, she is married to someone else. Nvidia has not responded to requests for comment.

Quellen:Tom's Hardware

Forschung

Baillie Gifford Investment Expert: Excess Returns in Healthcare AI Depend on a Few Companies Solving Hard Problems, Not Market Size

Wesentlich
2026-11-04

The investment logic of the Scottish Mortgage Investment Trust is not built on the common-sense notion that 'the healthcare market is large.' In arguments published in November 2026, Baillie Gifford investment specialist Hamish Maxwell explicitly stated that what truly attracts them is a few companies attempting to solve extremely difficult problems with new methods, thereby potentially gaining disproportionate returns. He wrote, 'The attraction is not just that healthcare is a large market, but that we see a few companies trying to solve very hard problems in novel ways, with potentially outsized rewards.' This judgment forms the core premise of the entire argument: the investment value of healthcare AI comes from the intersection of problem difficulty and method novelty, rather than track capacity.

When explaining how value is specifically created, Maxwell provided two types of mechanisms. The first is the data moat. Taking Tempus AI as an example, the company organizes and analyzes clinical and molecular data to help doctors make better decisions and researchers develop better therapies. Maxwell emphasized that in healthcare, data only constitutes a true competitive barrier when it is 'unique and difficult for competitors to replicate or buy.' Tempus claims connections with more than half of US oncologists and over 65% of US academic medical centers; Maxwell's reasoning is that the more decisions its data influences, the more valuable and harder to replace it becomes. This logic defines data advantage as an asset that deepens with use rather than decays over time.

The second mechanism is platform economics. Taking Moderna as an example, its most well-known product is the COVID-19 vaccine, but its broader ambition is to use messenger RNA (mRNA) as a programmable drug. In 2026, its personalized melanoma therapy (developed in collaboration with Merck and used in combination with Keytruda) became the first mRNA-based cancer therapy to succeed in Phase 3 clinical trials. Maxwell's reasoning is that drug development is typically slow, expensive, and highly specific, but a platform that can generate multiple products from the same technological foundation offers 'repeated innovation,' potentially changing the structure of economics and the scope of what can be created. Insulet’s Omnipod insulin delivery system demonstrates another dimension: for patients with chronic conditions, better devices can change lives, generating recurring revenue, patient loyalty, and long-term growth.

However, Maxwell inserted a key and frank limitation in his argument: one should not assume medicine will keep pace simply because AI is advancing rapidly. He wrote that AI technology relies on vast, diverse datasets and rapid feedback; in contrast, the clinical data relied upon by drug trials is slow and expensive to generate, designed to satisfy regulators rather than train algorithms. It may take years for medical regulators and clinical practice to catch up with genuine breakthroughs. This bottleneck means the path from innovation to impact may be 'slow and painful,' and 'few companies get to reshape the future of human health.' This is not an abstract warning, but a structural constraint directly affecting the time scale of investment returns.

Quellen:bailliegifford.com

Forschung
Weiter →
Anmelden / Registrieren
LesenForschungRadarAnlageframework

Therefore, the solution is not a larger data warehouse, but establishing an enterprise-level context layer. This utilizes controlled and consistent mechanisms to discover, connect, and retrieve data across systems and locations, enabling AI to avoid handing every data source to a central repository for processing. Metadata and vector information can assist AI in locating and inferring records restricted by regulations, sovereignty, and ownership that cannot be moved arbitrarily.

Chawla advises enterprises to answer four questions before executing major AI infrastructure investments: Which datasets are restricted from movement by regulations or contracts? Who holds governance rights over each dataset, and what happens when copies exist elsewhere? How do copies stay synchronized as source systems evolve? Can intelligent processing be deployed to the location where data currently resides, or does the architecture require all data to be centralized?

It should be noted that this framework comes from remarks by an executive at an infrastructure vendor, carrying a clear commercial stance. The 'context layer' described by Chawla remains conceptual, without providing specific implementation cases or quantitative effect comparisons. While 'data gravity' serves as an intuitively reasonable explanatory mechanism, its boundaries—such as at what scale or under which industry constraints this mechanism truly constitutes a bottleneck—have not yet been independently verified.

However, attributing the gap from pilot to scale-up to a named architectural constraint, rather than the vague 'data unready,' indeed provides practitioners with a more specific diagnostic starting point. The enterprises that stand to benefit most from AI in the future may be those that support the context required by AI agents when designing data architectures, without needing to move or centrally aggregate data across the entire organization.

In Alibaba Cloud's own practice, Zheng Tian's team chose a slightly different path: using Channel partitioning instead of a single Router, using non-resident Coordinators instead of persistent sessions, and using Quests (assignable, trackable, and verifiable work records) instead of Tickets. They discovered that Agents waking each other up is the most thorny pitfall in production environments—an organizational roll-call task that should have taken 3 executions actually ran more than 10 times. The root cause was that the platform placed 'being mentioned' and 'being assigned a task' into the same flow. The fix was to strictly distinguish semantics: '@' is communication, while creating a task is delegation. This conclusion aligns with DSH's: hard platform constraints are more reliable than prompts.

Finally, Zheng Tian places two trends side by side: architectures are converging, while Agents are returning to traditional engineering. As Agents shift from single conversations to long-running processes, they must face old problems like process crashes, machine offline states, duplicate messages, and side effects that cannot execute twice. In both DSH's code and his own, he increasingly sees classic database and distributed-system technologies—write-ahead logs, Outbox patterns, leases, and replayable event records. His summary is: the model is the CPU, and the Harness is the operating system and database.

Maxwell ultimately framed this argument as a belief in 'continuous reinvention' rather than a bet on a single technology, disease area, or regulatory outcome. Holdings in the healthcare innovation theme account for approximately 7% of the Scottish Mortgage portfolio. He acknowledged that 'some innovators will fail, others will disappoint,' but believed the returns are unusually alluring—supporting those companies that might create significant economic value and contribute to a world of longer, healthier lives. This conclusion preserves the conditional nature of the argument: it is not a prediction of success, but an argument that under specific mechanisms (deepening data moats, platform repeated innovation, eventual completion of clinical translation), the success of a few companies can generate returns far exceeding their current scale.

Notably, this argument comes from Baillie Gifford, which manages the Scottish Mortgage Investment Trust, a fund known for holding high-growth companies for the long term, having held genomic sequencing pioneer Illumina for a long period (2011 to 2023) and supporting it through a hostile takeover bid. Maxwell used this experience as context for understanding 'what better biological insight can deliver.' The article did not disclose specific holding costs, current valuation multiples, or expected return figures; its value lies in mechanism explanation and honest acknowledgment of bottlenecks, rather than verifiable performance records.