π₯ Top Story
The 24-hour experiment that helped Anthropic find its identity
This The New Stack piece examines how Anthropic came to treat evaluations, not product requirement documents (PRDs), as the central planning tool for frontier AI development. Traditional software teams write a PRD to agree on what to build, but the article argues that for frontier AI, defining the desired behavior through an eval is a better fit than freezing requirements in a document. The "24-hour experiment" in the title refers to a case where this approach reshaped not just a single project but the team's working identity. The underlying idea is that when model capabilities shift quickly, agreeing upfront on "how we'll measure success" is more practical than locking in a fixed spec. The piece frames this as evidence that AI product development is diverging from the traditional plan-then-build software engineering model.
π‘ Why it matters: Replacing requirement docs with evals as the source of truth gives teams a way to agree on "done" even as underlying model capabilities keep shifting.
π Read more Β· The New Stack
Kubernetes & Cloud Native
Federating clusters for zero-downtime Kubernetes
CNCF
This CNCF blog post highlights a situation multi-region Kubernetes operators run into often: when an entire cluster in one region disappears, having an identical copy of the service running two regions over is worthless if nothing is actually wired to route traffic there. The proposed fix is cluster federation β managing multiple Kubernetes clusters as a single logical unit rather than isolated islands. The core idea is that federation lets service discovery and traffic routing automatically fail over to the surviving region's identical workload when an entire region goes down. The piece draws a clear distinction between simply having replicas spread across regions (multi-region deployment) and actually achieving seamless failover when disaster strikes (federation). For SREs and platform engineers designing multi-region Kubernetes for high availability, it's a reminder that the federation wiring itself is what turns redundancy into real resilience.
π‘ Redundant replicas across regions don't guarantee zero downtime by themselves β clusters need to be actually federated so traffic fails over automatically when a region goes dark.
AI & ML
NVIDIA Cosmos-H-Dreams: Bringing Real-Time Generative Simulation to Surgical Robotics
Hugging Face
This Hugging Face blog post introduces Cosmos-H-Dreams, a new entry in NVIDIA's Cosmos model family. As the title states, it's focused on bringing real-time generative simulation to surgical robotics. NVIDIA's Cosmos line is known as a family of world foundation models used to generate physically plausible video and scenarios for training robots and other "physical AI" systems, and this release reads as an attempt to extend that scope into medical and surgical robotics specifically. Applying generative simulation to a domain with as little error tolerance as surgery raises the practical need for safety validation processes alongside the realism of the simulated data itself. For engineers working on medical robotics or simulation-based training pipelines, it's worth watching how the Cosmos family is branching from general-purpose robotics into specialized, high-stakes domains.
π‘ As generative world models extend into low-error-tolerance domains like surgical robotics, simulation realism alone isn't enough β safety validation processes need to scale alongside it.
How AI is expanding what people do at work
OpenAI
This OpenAI research post analyzes real usage patterns among ChatGPT users to examine how AI is expanding the boundaries of what people do at work. Its central claim is that AI isn't simply replacing existing tasks β it's enabling workers to take on tasks that used to fall outside their formal role. The research reportedly points to examples like marketers doing basic data analysis themselves, or non-engineers writing simple code, illustrating how role boundaries are blurring. This stands out because it offers a different framing than the usual "job replacement" narrative around AI adoption: role expansion rather than substitution. For organizations, the implication is that job descriptions may need to shift from narrow, fixed scopes toward more flexible role definitions that assume AI-assisted work as the norm.
π‘ If AI is expanding roles rather than just automating tasks away, organizations may need to redesign job descriptions around flexible, AI-assisted scopes instead of narrow fixed responsibilities.
Cloud Updates
Announcing general availability of SAP Business Data Cloud Connect for BigQuery
Google Cloud
Google Cloud announced the general availability of SAP Business Data Cloud Connect for BigQuery, a joint offering built with SAP. The service targets a common pain point in data replication: traditional pipelines often struggle to keep analytics data as fresh as modern workflows require, creating a lag between source systems and the data available for analysis. By connecting SAP data directly into BigQuery, the GA release aims to let organizations query near-current SAP data without building and maintaining a separate custom replication pipeline. This is particularly relevant for enterprises that run SAP as a system of record and want their analytics environment to reflect operational data with less delay. The move continues Google Cloud's push to deepen BigQuery's integration with major enterprise systems of record beyond generic ETL connectors.
π‘ Direct SAP-to-BigQuery integration cuts both the data-freshness lag and the pipeline maintenance burden for enterprises that treat SAP as their system of record.
Cyber Snapshot Report: Go beyond the toolchain and build enterprise resilience
Google Cloud
This Google Cloud blog post introduces a "Cyber Snapshot Report" built on Mandiant's frontline observations, arguing that even as attacks grow machine-speed fast, most successful breaches still trace back to fundamental human and organizational failures rather than toolchain gaps. The piece ties this pattern directly to findings from Mandiant's M-Trends 2026 report. The core message is that no amount of modern security tooling compensates for broken basics like process discipline, access management, and cross-team coordination. The report's recommendation is that organizations look beyond the toolchain and invest in building resilience across the whole organization, not just accumulating more security products. This aligns with a broader industry shift toward treating security spend as a people-and-process problem rather than purely a tooling problem.
π‘ Since most breaches trace back to process and access-control gaps rather than missing tools, security budgets are better spent hardening operational basics before adding another point product.
Modernizing the skies: NOAA and Google Cloud collaborate to advance weather forecasting
Google Cloud
This Google Cloud blog post covers a collaboration with NOAA (the National Oceanic and Atmospheric Administration) aimed at modernizing how the agency understands and predicts atmospheric patterns. The stated goal is to make weather forecasting β something that affects people's daily lives β more accurate and timely by upgrading NOAA's observation and computing infrastructure. The core of the partnership appears to be using Google Cloud's large-scale compute and data infrastructure to run more sophisticated weather models faster than before. It's an example of a public-sector agency adopting cloud infrastructure to scale up the size and speed of scientific computation, reflecting how cloud and AI capacity are increasingly embedded in climate and weather forecasting work. It's a useful reference case for any organization with large-scale simulation workloads considering a move to cloud infrastructure.
π‘ Moving large-scale public-sector simulation workloads to the cloud is a template for organizations that need elastic compute capacity beyond what on-prem infrastructure can practically support.
Weβre open-sourcing our privacy proxy CLI
Cloudflare
Cloudflare's blog announced that it has open-sourced pvcli, its privacy proxy command-line tool. Described as a curl-like utility, pvcli is designed to simplify testing complex privacy-preserving protocols such as OHTTP (Oblivious HTTP). Protocols in the OHTTP family split a request across multiple relay layers to hide the sender's identifying information, which makes them notoriously fiddly to implement and debug β pvcli's goal is to let developers exercise that handshake and encryption layering with simple, curl-style commands instead of hand-rolling test clients. By open-sourcing the tool, Cloudflare appears to be trying to widen the ecosystem of developers and researchers who can easily experiment with and verify interoperability of these privacy protocols. For engineers evaluating privacy-enhancing technologies, it adds a practical, ready-made tool for observing protocol behavior directly rather than building test tooling from scratch.
π‘ A curl-like open-source tool for testing OHTTP-style protocols significantly lowers the barrier to verifying protocol behavior before adopting privacy-enhancing tech in production.
Strengthening the open source defense layer: Red Hat joins NVIDIA in the Open Secure AI Alliance
Red Hat
This Red Hat blog post announces that the company has joined NVIDIA's Open Secure AI Alliance. It starts from the premise that as AI capabilities advance, the security landscape shifts in real time, and no single company can keep up with that pace acting alone. The alliance is presented as a cross-company effort, built on the open source ecosystem, to collectively build up AI security capabilities. For Red Hat, the move is framed as strengthening the "open source defense layer" that underpins trust in the software supply chain. It's an example of the industry trying to address AI infrastructure security through shared standards and collaboration rather than leaving it to individual vendors.
π‘ Since no single vendor can keep pace with how fast AI security threats evolve, participation in open, cross-industry alliances is becoming a real factor in long-term AI infrastructure defense.
How leading companies are turning AI vision into business value
Red Hat
This Red Hat blog post draws on the author's experience speaking with hundreds of customers a year to explore why some companies succeed at turning AI vision into actual business value while others don't. It opens by noting that customers span the full spectrum, from leading-edge early adopters to highly risk-averse, conservative organizations. The core observation seems to be that success hinges less on declaring an AI strategy and more on the execution discipline of tying that strategy to real operational processes and performance metrics. The piece leans more on qualitative, leadership-level insight than on specific case studies or figures, making it more relevant to executives shaping AI strategy than to engineering teams in the weeds. For anyone driving AI initiatives inside an organization, it's a reminder to focus on closing the gap between stated vision and measurable, realized value.
π‘ AI initiatives succeed or fail based on whether the strategy is actually wired into operational processes and metrics β a vision statement without an execution plan behind it is a warning sign.
Building the future: Core concepts of Red Hatβs agentic software development life cycle
Red Hat
This Red Hat blog post lays out the core concepts behind an "agentic software development life cycle" (SDLC), arguing that when integrating AI into software development, the methodology matters just as much as the underlying models. Red Hat's position is that AI's real value only shows up when it's embedded into enterprise workflows that are transparent, scalable, and reliable. That appears to mean weaving AI agents naturally into each stage of the traditional SDLC β requirements, design, implementation, testing, deployment β while keeping enough transparency that humans can inspect the reasoning behind each stage's output. It reads as a methodological approach to making agents participate in the development lifecycle in a repeatable, auditable way, rather than treating them as one-off code generation tools. It's a relevant reference for platform teams trying to design AI agent adoption in enterprise settings with governance in mind.
π‘ Safely folding AI agents into the SDLC requires a methodology for transparently tracking and verifying each stage's decisions β before model choice, not after.
DevOps & Infrastructure
Moonshot opens Kimi K3 weights β but few can run it
The New Stack
According to The New Stack, Chinese AI startup Moonshot AI has released the open weights for its latest model, Kimi K3, on Hugging Face, making them freely downloadable for developers. As the headline points out, however, publishing the weights doesn't mean most people can actually run the model. Modern large language models tend to demand proportionally larger GPU memory and serving infrastructure as their parameter count grows, so in practice only organizations with substantial compute clusters can self-host the largest releases. The piece frames this as a gap between the licensing sense of "openness" and the practical accessibility of actually using the model. It's presented as part of a broader pattern in which open-weight releases keep arriving, but downloadability alone doesn't guarantee real-world usability for most teams.
π‘ Even open-weight models can require serious GPU and infrastructure investment to self-host, so teams should size the inference cost and hardware footprint before committing to one.
The harness is all you need (mostly)
GitHub
This GitHub blog post argues for building a practical development workflow around GitHub Copilot rather than constantly chasing whichever new AI tool launches next. It walks through using a single "harness" β a consistent workflow wrapper β to handle prototyping, planning, implementation, and code review in one coherent loop. The title's "the harness is all you need (mostly)" reflects the piece's central claim: the workflow design around a tool tends to matter more for productivity than the specific model or tool being used at any given moment. Rather than paying a re-learning cost every time developers switch tools, the article suggests treating models as swappable components sitting inside a stable harness. For engineering leaders trying to set team-wide standards amid a flood of new AI coding tools, this offers a concrete way to think about workflow durability over tool churn.
π‘ Standardizing the harness around AI coding tools β rather than the tools themselves β keeps a team's workflow stable even as the underlying models keep changing.
Microsoft is racing to make OpenAI optional
The New Stack
This The New Stack article looks at Microsoft's push to reduce its reliance on OpenAI. It opens with the observation that the AI landscape is moving so fast that Microsoft CEO Satya Nadella has taken to posting lengthy threads on X (Twitter) to keep up. The piece connects this pace to a broader industry read: even as Microsoft has deeply integrated OpenAI's models across products like Copilot, it's simultaneously building up its own in-house AI capabilities to reduce strategic dependence on any single model supplier. The OpenAI partnership remains a central pillar, but the underlying logic is that securing multiple model sources is simply sound risk management. It's presented as part of a wider pattern of cloud and enterprise software vendors diversifying their AI model sourcing rather than betting everything on one provider.
π‘ Concentrating core AI features on a single model vendor creates exposure to pricing and policy shifts, which is why multi-model sourcing is becoming a risk-management practice for enterprise architects.
GitHub Copilot app for Beginners: Getting started
GitHub
This GitHub blog post is a beginner-focused guide to the GitHub Copilot app. It walks through the app's core capabilities step by step: starting new projects, working alongside AI agents, and exploring the "canvas" feature, all aimed at helping developers pick up the tool without a complicated setup process. The intent is clearly onboarding β lowering the entry barrier for developers who are new to Copilot as AI coding tools keep multiplying. For engineering managers rolling out Copilot to a team for the first time, this doubles as ready-made onboarding material. It's a practical, low-friction reference rather than a deep technical dive.
π‘ Having an official beginner's guide on hand shortens the learning curve when rolling Copilot out to a team, reducing the usual gap between early adopters and everyone else.
Smarter onboarding and planning with Grafana Assistant: How to ensure observability is baked in from the start
Grafana
This Grafana blog post opens with a familiar scenario: a feature is basically done, but "add monitoring" is still sitting untouched at the bottom of the ticket. The central pitch is that Grafana Assistant can help teams bring observability into the onboarding and planning stage instead of bolting it on at the last minute. By suggesting which metrics, logs, and alerts should be designed upfront, the AI assistant reframes observability as a built-in requirement rather than an afterthought squeezed in before launch. This fits into the broader "observability by design" movement, which favors thinking about visibility during design rather than reactive monitoring after incidents. It's a useful reference for teams looking to bake an observability checklist into their early sprint planning.
π‘ Designing monitoring in at the planning stage with an assistant, rather than bolting it on late, cuts down on the all-too-common situation of discovering missing metrics only after an incident.
Explore what's next in agentic operations: Introducing AI Week
Grafana
This Grafana blog post introduces "AI Week," a content series positioned around the rise of "agentic operations." It starts from the observation that observability has traditionally been bolted on after code reaches production, but that reactive approach no longer holds up once AI agents are directly performing operational work. AI Week appears to be a set of sessions and content exploring where observability and automation need to head in an agent-driven operations environment. The underlying argument connects to a broader industry concern: as AI agents get more deeply involved in running infrastructure, monitoring tooling itself needs to be redesigned around observing agents, not just traditional services. For teams introducing AI agents into operational automation, it's a prompt to think ahead about how to track and verify agent behavior rather than treating it as a black box.
π‘ Once AI agents are actively operating infrastructure, after-the-fact monitoring isn't enough β teams need to design real-time observability into agent behavior itself, upfront.
Claude Opus 5 on GitLab: Reasoning built for the hard tasks
GitLab
This GitLab blog post covers the availability of Anthropic's Claude Opus 5 on the GitLab Duo Agent Platform. It opens by contrasting the cost of mistakes: a slip on a routine task might cost a minute, but an error in a large refactor or a debugging trail spanning months of commit history can compound silently across hundreds of exchanges and cost far more. In that context, Claude Opus 5 is positioned as the model suited to harder reasoning tasks β the kind of complex refactors and long-horizon debugging where careful judgment matters most. Integrating it into the GitLab Duo Agent Platform gives developers the option to delegate exactly those higher-stakes tasks to an agent from within the platform. It also reflects a broader trend of matching model choice to task difficulty β lighter, faster models for routine work, and stronger reasoning models for complex, high-risk work.
π‘ Routing complex refactors and long-running debugging work to a reasoning-focused model, while keeping lighter models for routine tasks, reduces the compounding risk of silent errors across a development pipeline.
What Is AI Pentesting and How Does It Works?
Snyk
This Snyk blog post explains what AI pentesting is and how it works. The core idea is using reasoning-capable AI models to continuously find and validate flaws that traditional static and dynamic vulnerability scanners tend to miss. It specifically calls out broken authorization and business-logic abuse as categories where this approach shines β the kinds of vulnerabilities that pattern-matching scanners struggle to catch. Where conventional scanners mostly hunt for known vulnerability signatures, AI pentesting instead reasons about an application's context and logic to judge whether a given flow is actually exploitable. For security teams looking to move beyond the traditional once-or-twice-a-year penetration test toward continuous vulnerability discovery and validation, this offers a practical starting point.
π‘ Using AI pentesting to continuously probe for authorization and business-logic flaws β the blind spots of traditional scanners β helps close the gap left by annual or biannual pentest cycles.
This digest was collected from RSS feeds and summarized by AI (Claude). See the original links for full details.